fix: bypass gloo DDP for Windows single-GPU training

On Windows with a single GPU, dist.init_process_group() using the
gloo backend frequently fails with 'unsupported gloo device', caused
by virtual network adapters (VPN, VMware, Hyper-V, etc.).

Changes:
- s2_train.py: skip dist.init_process_group() on Windows single-GPU;
  add DummyDDP wrapper to maintain .module interface compatibility
- s1_train.py: set USE_LIBUV=0 to avoid socket conflicts;
  use strategy='auto' for single-GPU (bypasses gloo entirely),
  DDPStrategy only activated for multi-GPU setups
- utils.py, bucket_sampler.py: related compatibility adjustments

Tested on: Windows 11, single NVIDIA GPU, Python 3.10, PyTorch 2.5

GPT_SoVITS/AR/data/bucket_sampler.py

@@ -39,12 +39,12 @@ class DistributedBucketSampler(Sampler[T_co]):
         if num_replicas is None:
             if not dist.is_available():
                 raise RuntimeError("Requires distributed package to be available")
-            num_replicas = dist.get_world_size() if torch.cuda.is_available() else 1
+            num_replicas = dist.get_world_size() if dist.is_initialized() else 1
         if rank is None:
             if not dist.is_available():
                 raise RuntimeError("Requires distributed package to be available")
-            rank = dist.get_rank() if torch.cuda.is_available() else 0
+            rank = dist.get_rank() if dist.is_initialized() else 0
-            if torch.cuda.is_available():
+            if torch.cuda.is_available() and dist.is_initialized():
                 torch.cuda.set_device(rank)
         if rank >= num_replicas or rank < 0:
             raise ValueError("Invalid rank {}, rank should be in the interval [0, {}]".format(rank, num_replicas - 1))

GPT_SoVITS/s1_train.py

@@ -118,7 +118,7 @@ def main(args):
         benchmark=False,
         fast_dev_run=False,
         strategy=DDPStrategy(process_group_backend="nccl" if platform.system() != "Windows" else "gloo")
-        if torch.cuda.is_available()
+        if torch.cuda.is_available() and torch.cuda.device_count() > 1
         else "auto",
         precision=config["train"]["precision"],
         logger=logger,

GPT_SoVITS/s2_train.py

@@ -77,6 +77,7 @@ def run(rank, n_gpus, hps):
         writer = SummaryWriter(log_dir=hps.s2_ckpt_dir)
         writer_eval = SummaryWriter(log_dir=os.path.join(hps.s2_ckpt_dir, "eval"))
 
+    if not (os.name == "nt" and n_gpus == 1):
         dist.init_process_group(
             backend="gloo" if os.name == "nt" or not torch.cuda.is_available() else "nccl",
             init_method="env://?use_libuv=False",
@@ -197,6 +198,16 @@ def run(rank, n_gpus, hps):
         eps=hps.train.eps,
     )
     if torch.cuda.is_available():
+        if os.name == "nt" and n_gpus == 1:
+            class DummyDDP(torch.nn.Module):
+                def __init__(self, module):
+                    super().__init__()
+                    self.module = module
+                def forward(self, *args, **kwargs):
+                    return self.module(*args, **kwargs)
+            net_g = DummyDDP(net_g)
+            net_d = DummyDDP(net_d)
+        else:
             net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
             net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
     else:

GPT_SoVITS/utils.py

@@ -64,12 +64,25 @@ import shutil
 from time import time as ttime
 
 
+import time
+
 def my_save(fea, path):  #####fix issue: torch.save doesn't support chinese path
     dir = os.path.dirname(path)
     name = os.path.basename(path)
-    tmp_path = "%s.pth" % (ttime())
+    tmp_path = "%s.pth" % (time.time())
     torch.save(fea, tmp_path)
-    shutil.move(tmp_path, "%s/%s" % (dir, name))
+    target_path = "%s/%s" % (dir, name)
+    try:
+        shutil.move(tmp_path, target_path)
+    except Exception as e:
+        print(f"Move failed with error {e}, retrying via copy and delete...")
+        if os.path.exists(target_path):
+            try:
+                os.remove(target_path)
+            except:
+                pass
+        shutil.copyfile(tmp_path, target_path)
+        os.remove(tmp_path)
 
 
 def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):