Add LM Backend selector (PT vs VLLM) to reduce VRAM usage

PT backend uses ~1.6 GB VRAM vs ~9.2 GB for VLLM, making it accessible
on lower-end GPUs. Adds dropdown in Advanced Settings defaulting to PT.

App.tsx

@@ -596,6 +596,7 @@ export default function App() {
         lmTopK: params.lmTopK,
         lmTopP: params.lmTopP,
         lmNegativePrompt: params.lmNegativePrompt,
+        lmBackend: params.lmBackend,
         referenceAudioUrl: params.referenceAudioUrl,
         sourceAudioUrl: params.sourceAudioUrl,
         audioCodes: params.audioCodes,

components/CreatePanel.tsx

@@ -133,6 +133,7 @@ export const CreatePanel: React.FC<CreatePanelProps> = ({ onGenerate, isGenerati
   const [audioFormat, setAudioFormat] = useState<'mp3' | 'flac'>('mp3');
   const [inferenceSteps, setInferenceSteps] = useState(8);
   const [inferMethod, setInferMethod] = useState<'ode' | 'sde'>('ode');
+  const [lmBackend, setLmBackend] = useState<'pt' | 'vllm'>('pt');
   const [shift, setShift] = useState(3.0);
 
   // LM Parameters (under Expert)
@@ -535,6 +536,7 @@ export const CreatePanel: React.FC<CreatePanelProps> = ({ onGenerate, isGenerati
         thinking,
         audioFormat,
         inferMethod,
+        lmBackend,
         shift,
         lmTemperature,
         lmCfgScale,
@@ -1298,6 +1300,20 @@ export const CreatePanel: React.FC<CreatePanelProps> = ({ onGenerate, isGenerati
               </div>
             </div>
 
+            {/* LM Backend */}
+            <div className="space-y-1.5">
+              <label className="text-xs font-medium text-zinc-600 dark:text-zinc-400">LM Backend</label>
+              <select
+                value={lmBackend}
+                onChange={(e) => setLmBackend(e.target.value as 'pt' | 'vllm')}
+                className="w-full bg-zinc-50 dark:bg-black/20 border border-zinc-200 dark:border-white/10 rounded-lg px-2 py-1.5 text-xs text-zinc-900 dark:text-white focus:outline-none"
+              >
+                <option value="pt">PT (~1.6 GB VRAM)</option>
+                <option value="vllm">VLLM (~9.2 GB VRAM)</option>
+              </select>
+              <p className="text-[10px] text-zinc-500">PT uses less VRAM, VLLM may be faster on powerful GPUs</p>
+            </div>
+
             {/* Seed */}
             <div className="space-y-2">
               <div className="flex items-center justify-between">

server/src/routes/generate.ts

@@ -90,6 +90,7 @@ interface GenerateBody {
   lmTopK?: number;
   lmTopP?: number;
   lmNegativePrompt?: string;
+  lmBackend?: 'pt' | 'vllm';
 
   // Expert Parameters
   referenceAudioUrl?: string;
@@ -189,6 +190,7 @@ router.post('/', authMiddleware, async (req: AuthenticatedRequest, res: Response
       lmTopK,
       lmTopP,
       lmNegativePrompt,
+      lmBackend,
       referenceAudioUrl,
       sourceAudioUrl,
       audioCodes,
@@ -252,6 +254,7 @@ router.post('/', authMiddleware, async (req: AuthenticatedRequest, res: Response
       lmTopK,
       lmTopP,
       lmNegativePrompt,
+      lmBackend,
       referenceAudioUrl,
       sourceAudioUrl,
       audioCodes,

server/src/services/acestep.ts

@@ -137,6 +137,7 @@ async function submitToApi(params: GenerationParams): Promise<{ taskId: string }
     use_cot_caption: false, // Explicitly disable CoT features that require LLM
     use_cot_language: false, // Explicitly disable CoT features that require LLM
     use_cot_metas: false, // Explicitly disable CoT features that require LLM
+    lm_backend: params.lmBackend || 'pt',
   };
 
   if (params.bpm && params.bpm > 0) body.bpm = params.bpm;
@@ -150,7 +151,12 @@ async function submitToApi(params: GenerationParams): Promise<{ taskId: string }
   if (params.audioCodes) body.audio_code_string = params.audioCodes;
   if (params.repaintingStart !== undefined && params.repaintingStart > 0) body.repainting_start = params.repaintingStart;
   if (params.repaintingEnd !== undefined && params.repaintingEnd > 0) body.repainting_end = params.repaintingEnd;
-  if (params.audioCoverStrength !== undefined && params.audioCoverStrength !== 1.0) body.audio_cover_strength = params.audioCoverStrength;
+  // Always send audio_cover_strength for cover/repaint tasks, otherwise only when not default
+  if (params.taskType === 'cover' || params.taskType === 'repaint' || params.sourceAudioUrl) {
+    body.audio_cover_strength = params.audioCoverStrength ?? 1.0;
+  } else if (params.audioCoverStrength !== undefined && params.audioCoverStrength !== 1.0) {
+    body.audio_cover_strength = params.audioCoverStrength;
+  }
   if (params.instruction) body.instruction = params.instruction;
   // LLM and CoT parameters only sent when thinking mode is enabled
   if (params.thinking) {
@@ -340,6 +346,7 @@ export interface GenerationParams {
   lmTopK?: number;
   lmTopP?: number;
   lmNegativePrompt?: string;
+  lmBackend?: 'pt' | 'vllm';
 
   // Expert Parameters
   referenceAudioUrl?: string;
@@ -603,7 +610,12 @@ async function processGeneration(
     if (params.audioCodes) args.push('--audio-codes', params.audioCodes);
     if (params.repaintingStart !== undefined && params.repaintingStart > 0) args.push('--repainting-start', String(params.repaintingStart));
     if (params.repaintingEnd !== undefined && params.repaintingEnd > 0) args.push('--repainting-end', String(params.repaintingEnd));
-    if (params.audioCoverStrength !== undefined && params.audioCoverStrength !== 1.0) args.push('--audio-cover-strength', String(params.audioCoverStrength));
+    // Always send audio_cover_strength for cover/repaint tasks, otherwise only when not default
+    if (params.taskType === 'cover' || params.taskType === 'repaint' || params.sourceAudioUrl) {
+      args.push('--audio-cover-strength', String(params.audioCoverStrength ?? 1.0));
+    } else if (params.audioCoverStrength !== undefined && params.audioCoverStrength !== 1.0) {
+      args.push('--audio-cover-strength', String(params.audioCoverStrength));
+    }
     if (params.instruction) args.push('--instruction', params.instruction);
     if (params.thinking) args.push('--thinking');
     if (params.lmTemperature !== undefined) args.push('--lm-temperature', String(params.lmTemperature));
@@ -611,6 +623,7 @@ async function processGeneration(
     if (params.lmTopK !== undefined && params.lmTopK > 0) args.push('--lm-top-k', String(params.lmTopK));
     if (params.lmTopP !== undefined) args.push('--lm-top-p', String(params.lmTopP));
     if (params.lmNegativePrompt) args.push('--lm-negative-prompt', params.lmNegativePrompt);
+    if (params.lmBackend) args.push('--lm-backend', params.lmBackend);
     if (params.useCotMetas === false) args.push('--no-cot-metas');
     if (params.useCotCaption === false) args.push('--no-cot-caption');
     if (params.useCotLanguage === false) args.push('--no-cot-language');

services/api.ts

@@ -232,6 +232,7 @@ export interface GenerationParams {
   lmTopK?: number;
   lmTopP?: number;
   lmNegativePrompt?: string;
+  lmBackend?: 'pt' | 'vllm';
 
   // Expert Parameters
   referenceAudioUrl?: string;

types.ts

@@ -83,6 +83,7 @@ export interface GenerationParams {
   lmTopK: number;
   lmTopP: number;
   lmNegativePrompt: string;
+  lmBackend?: 'pt' | 'vllm';
 
   // Expert Parameters
   referenceAudioUrl?: string;