Use file-system torch sharing for training

app.py

@@ -255,6 +255,20 @@ def run_cmd(command, cwd=None, env=None):
         raise RuntimeError(f"命令失败 (exit={code}): {' '.join(command)}")
 
 
+def python_script_command(script_path, *script_args, file_system_sharing=False):
+    if not file_system_sharing:
+        return [sys.executable, "-s", script_path, *[str(arg) for arg in script_args]]
+    launcher = (
+        "import runpy, sys;"
+        "import torch.multiprocessing as mp;"
+        "mp.set_sharing_strategy('file_system');"
+        "script = sys.argv[1];"
+        "sys.argv = sys.argv[1:];"
+        "runpy.run_path(script, run_name='__main__')"
+    )
+    return [sys.executable, "-s", "-c", launcher, script_path, *[str(arg) for arg in script_args]]
+
+
 def has_transformers_model(path: Path):
     return path.exists() and (path / "config.json").exists() and (
         any(path.glob("*.bin")) or any(path.glob("*.safetensors"))
@@ -864,7 +878,7 @@ def prepare_data_steps(logs, version, live_path: Path | None = None):
             )
             env = build_process_env(ctx)
             for line in run_cmd(
-                [sys.executable, "-s", "GPT_SoVITS/prepare_datasets/1-get-text.py"],
+                python_script_command("GPT_SoVITS/prepare_datasets/1-get-text.py"),
                 cwd=GPT_SOVITS_DIR,
                 env=env,
             ):
@@ -875,7 +889,7 @@ def prepare_data_steps(logs, version, live_path: Path | None = None):
             part_text.replace(ctx.text_path)
             yield push(logs, "✅ 文本分词与 BERT 特征提取完成。", live_path)
             for line in run_cmd(
-                [sys.executable, "-s", "GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py"],
+                python_script_command("GPT_SoVITS/prepare_datasets/2-get-hubert-wav32k.py"),
                 cwd=GPT_SOVITS_DIR,
                 env=env,
             ):
@@ -883,14 +897,14 @@ def prepare_data_steps(logs, version, live_path: Path | None = None):
             yield push(logs, "✅ CN-HuBERT 特征与 32k wav 已生成。", live_path)
             if ctx.spec.uses_sv:
                 for line in run_cmd(
-                    [sys.executable, "-s", "GPT_SoVITS/prepare_datasets/2-get-sv.py"],
+                    python_script_command("GPT_SoVITS/prepare_datasets/2-get-sv.py"),
                     cwd=GPT_SOVITS_DIR,
                     env=env,
                 ):
                     yield push(logs, line, live_path)
                 yield push(logs, "✅ speaker embedding 特征已生成。", live_path)
             for line in run_cmd(
-                [sys.executable, "-s", "GPT_SoVITS/prepare_datasets/3-get-semantic.py"],
+                python_script_command("GPT_SoVITS/prepare_datasets/3-get-semantic.py"),
                 cwd=GPT_SOVITS_DIR,
                 env=env,
             ):
@@ -964,7 +978,12 @@ def start_training(version=DEFAULT_VERSION, epochs=2, batch_size=1, save_every_e
         env = build_process_env(ctx)
         yield push(logs, f"开始 SoVITS 训练（{ctx.spec.version}）...", ctx.sovits_live_log), None
         for line in run_cmd(
-            [sys.executable, "-s", "GPT_SoVITS/s2_train.py", "--config", str(config_path)],
+            python_script_command(
+                "GPT_SoVITS/s2_train.py",
+                "--config",
+                str(config_path),
+                file_system_sharing=True,
+            ),
             cwd=GPT_SOVITS_DIR,
             env=env,
         ):
@@ -994,7 +1013,12 @@ def start_gpt_training(version=DEFAULT_VERSION, epochs=1, batch_size=1, save_eve
         env = build_process_env(ctx)
         yield push(logs, f"开始 GPT 训练（{ctx.spec.version}）...", ctx.gpt_live_log), None
         for line in run_cmd(
-            [sys.executable, "-s", "GPT_SoVITS/s1_train.py", "--config_file", str(config_path)],
+            python_script_command(
+                "GPT_SoVITS/s1_train.py",
+                "--config_file",
+                str(config_path),
+                file_system_sharing=True,
+            ),
             cwd=GPT_SOVITS_DIR,
             env=env,
         ):