Remove YuE2 process memory fraction cap on the 5080.

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

scripts/yue2-worker.py

@@ -119,18 +119,24 @@ def main():
     patch_graph_attention(attention_backend)
     free_cuda(torch)
     memory_snapshot(torch, 'memory-before-load')
-    # 16GB card: cap budget so YuE2 leaves headroom; tiled VAE at 512 frames.
+    # Prefer tiled VAE; do not cap the process with set_per_process_memory_fraction.
     pipe_load = dict(
         device='cuda',
         backend=pipeline_backend,
-        memory_budget_gib=12,
         vae_core_frames=512,
         offload_ar=True,
         progress=False,
     )
     cot = 'full'
     pipe_kwargs = dict(style=style, lyrics=lyrics, cot=cot, seed=seed)
-    with YuE2Pipeline.from_pretrained(model, vae=vae, **pipe_load) as pipe:
+    # 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
+    try:
+        pipe_cm = YuE2Pipeline.from_pretrained(model, vae=vae, **pipe_load)
+    finally:
+        torch.cuda.set_per_process_memory_fraction = _set_fraction
+    with pipe_cm as pipe:
         memory_snapshot(torch, 'memory-loaded')
         emit(stage='plan', message='Planning melody and chords', progress=5, cot=cot)
         try: