Keep YuE2 jobs live on the bench and wire target length to semantic tokens.

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

scripts/yue2-worker.py

@@ -97,8 +97,10 @@ def main():
     root = Path(cli.root).resolve()
     output = Path(cli.request).resolve().parent
     duration = int(request.get('duration', 60))
-    if not 30 <= duration <= 150:
-        raise ValueError('Duration must be 30–150 seconds.')
+    if not 15 <= duration <= 150:
+        raise ValueError('Target length must be 15–150 seconds.')
+    # VAE downsampling_ratio 1920 @ 48 kHz → 25 semantic / latent frames per second.
+    max_tokens = max(200, min(9000, duration * 25))
     lyrics = normalize_lyrics(request['lyrics'])
     style = ' '.join(str(request['tags']).split())
     seed = int(request['seed'])
@@ -111,9 +113,11 @@ def main():
     import soundfile as sf
     attention_backend, pipeline_backend = resolve_attention_backend(torch)
     emit(stage='loading', progress=1, message='Loading YuE2', model=model, vae=vae,
-         duration=duration, attention_backend=attention_backend, pipeline_backend=pipeline_backend,
+         duration=duration, maxTokens=max_tokens, attention_backend=attention_backend,
+         pipeline_backend=pipeline_backend,
          cudaAllocConf=os.environ.get('PYTORCH_CUDA_ALLOC_CONF'))
     from yue2 import YuE2Pipeline
+    from yue2.protocol import Sampling
     if not torch.cuda.is_available():
         raise RuntimeError('YuE2 requires a CUDA GPU; CPU fallback is disabled.')
     patch_graph_attention(attention_backend)
@@ -129,6 +133,8 @@ def main():
     )
     cot = 'full'
     pipe_kwargs = dict(style=style, lyrics=lyrics, cot=cot, seed=seed)
+    # Cap semantic length from target seconds. Do not FFmpeg-trim after decode.
+    semantic_sampling = Sampling(max_tokens=max_tokens, min_tokens=min(200, max_tokens))
     # YuE2Pipeline.__init__ always calls set_per_process_memory_fraction; skip it on this 16GB host.
     _set_fraction = torch.cuda.set_per_process_memory_fraction
     torch.cuda.set_per_process_memory_fraction = lambda *args, **kwargs: None
@@ -153,8 +159,9 @@ def main():
             plan = pipe.plan(**pipe_kwargs)
         free_cuda(torch)
         memory_snapshot(torch, 'memory-after-plan')
-        emit(stage='semantic', message='Generating semantic tokens', progress=25, cot=cot)
-        semantic = pipe.generate_semantic(plan)
+        emit(stage='semantic', message='Generating semantic tokens', progress=25, cot=cot,
+             maxTokens=max_tokens, targetSeconds=duration)
+        semantic = pipe.generate_semantic(plan, sampling=semantic_sampling)
         free_cuda(torch)
         memory_snapshot(torch, 'memory-after-semantic')
         emit(stage='synthesize', message='Synthesizing acoustic latents', progress=55, cot=cot)