ZeroGPU: gloo pg init, drop cuda.set_device (emulation wall)

app.py

@@ -127,8 +127,20 @@ def load_pipeline() -> None:
     ]
     ff_args = ff_infer.build_parser().parse_args(argv)
 
-    # Same initialization order as infer.py main().
-    torch.cuda.set_device(0)
+    # Same initialization order as infer.py main(), adapted for ZeroGPU:
+    # - no torch.cuda.set_device: a ZeroGPU slice has one device, and the
+    #   low-level call is not intercepted by the CUDA emulation at startup.
+    # - pre-init a gloo (CPU) process group so their init_device_mesh("cuda")
+    #   does not try to auto-init NCCL; with world_size=1 and sp=1 the meshes
+    #   are 1x1, sp_group is None, and no collectives ever run.
+    import torch.distributed as dist
+    if not dist.is_initialized():
+        dist.init_process_group(
+            backend="gloo",
+            init_method="tcp://127.0.0.1:29511",
+            rank=0,
+            world_size=1,
+        )
     ip.initialize_parallel_state(sp=1)
     pipe, _ = ff_infer.build_pipeline(ff_args)
     print("FrameForge pipeline loaded.")