blackteay/GPT-SoVITS
1
0
Fork 0
mirror of https://github.com/RVC-Boss/GPT-SoVITS.git synced 2026-04-29 21:00:42 +08:00
Code Issues Packages Projects Releases Wiki Activity

[codex] Improve Windows single-GPU v3 LoRA training / 改进 Windows 单卡 v3 LoRA 训练流程 (#2767)

Browse Source 
* Improve Windows single-GPU v3 LoRA training

* Drop unrelated checkpoint helper change from PR

* Tighten PR scope to single-GPU training path fixes
This commit is contained in:
东云 2026-04-18 16:54:26 +08:00 committed by GitHub
parent ba8de9b760
commit 780383d5bd
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 28 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
GPT_SoVITS/s2_train_v3_lora.py
View File

@ -55,6 +55,10 @@ def main():
n_gpus = torch.cuda.device_count() n_gpus = torch.cuda.device_count()
else: else:
n_gpus = 1 n_gpus = 1
if n_gpus <= 1:
run(0, n_gpus, hps)
return
os.environ["MASTER_ADDR"] = "localhost" os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = str(randint(20000, 55555)) os.environ["MASTER_PORT"] = str(randint(20000, 55555))
@ -77,12 +81,14 @@ def run(rank, n_gpus, hps):
writer = SummaryWriter(log_dir=hps.s2_ckpt_dir) writer = SummaryWriter(log_dir=hps.s2_ckpt_dir)
writer_eval = SummaryWriter(log_dir=os.path.join(hps.s2_ckpt_dir, "eval")) writer_eval = SummaryWriter(log_dir=os.path.join(hps.s2_ckpt_dir, "eval"))
dist.init_process_group( use_ddp = n_gpus > 1
backend="gloo" if os.name == "nt" or not torch.cuda.is_available() else "nccl", if use_ddp:
init_method="env://?use_libuv=False", dist.init_process_group(
world_size=n_gpus, backend="gloo" if os.name == "nt" or not torch.cuda.is_available() else "nccl",
rank=rank, init_method="env://?use_libuv=False",
) world_size=n_gpus,
rank=rank,
)
torch.manual_seed(hps.train.seed) torch.manual_seed(hps.train.seed)
if torch.cuda.is_available(): if torch.cuda.is_available():
torch.cuda.set_device(rank) torch.cuda.set_device(rank)
@ -118,15 +124,20 @@ def run(rank, n_gpus, hps):
shuffle=True, shuffle=True,
) )
collate_fn = TextAudioSpeakerCollate() collate_fn = TextAudioSpeakerCollate()
train_loader = DataLoader( worker_count = 0 if os.name == "nt" and n_gpus <= 1 else min(2 if os.name == "nt" else 5, os.cpu_count() or 1)
train_dataset, loader_kwargs = dict(
num_workers=5, num_workers=worker_count,
shuffle=False, shuffle=False,
pin_memory=True, pin_memory=torch.cuda.is_available(),
collate_fn=collate_fn, collate_fn=collate_fn,
batch_sampler=train_sampler, batch_sampler=train_sampler,
persistent_workers=True, )
prefetch_factor=3, if worker_count > 0:
loader_kwargs["persistent_workers"] = True
loader_kwargs["prefetch_factor"] = 2 if os.name == "nt" else 3
train_loader = DataLoader(
train_dataset,
**loader_kwargs,
) )
save_root = "%s/logs_s2_%s_lora_%s" % (hps.data.exp_dir, hps.model.version, hps.train.lora_rank) save_root = "%s/logs_s2_%s_lora_%s" % (hps.data.exp_dir, hps.model.version, hps.train.lora_rank)
os.makedirs(save_root, exist_ok=True) os.makedirs(save_root, exist_ok=True)
@ -156,7 +167,9 @@ def run(rank, n_gpus, hps):
def model2cuda(net_g, rank): def model2cuda(net_g, rank):
if torch.cuda.is_available(): if torch.cuda.is_available():
net_g = DDP(net_g.cuda(rank), device_ids=[rank], find_unused_parameters=True) net_g = net_g.cuda(rank)
if use_ddp:
net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
else: else:
net_g = net_g.to(device) net_g = net_g.to(device)
return net_g return net_g
@ -242,6 +255,8 @@ def run(rank, n_gpus, hps):
None, None,
) )
scheduler_g.step() scheduler_g.step()
if use_ddp and dist.is_initialized():
dist.destroy_process_group()
print("training done") print("training done")