Update app.py

app.py

@@ -1,64 +1,9 @@
 import os
 import sys
+import shutil
 
 # ==========================================
-# 1. 核心补丁：全局搜索并摧毁 Flash Attention
-# ==========================================
-print("🛰️ 启动全盘扫描，寻找 GPU 毒瘤文件...")
-
-# 定义我们要写入的“CPU 伪装代码”
-# 这段代码会把所有对 GPU 模型的调用，转发给 CPU 模型
-CPU_HACK_CODE = """
-import sys
-import logging
-print("🛡️ [HACK] 成功拦截 FlashAttn 调用，强制重定向至 CPU 模型！")
-
-# 试图导入普通的 CPU 模型
-try:
-    # 尝试多种可能的路径
-    try:
-        from AR.models.t2s_model import Text2SemanticDecoder
-    except ImportError:
-        from GPT_SoVITS.AR.models.t2s_model import Text2SemanticDecoder
-except ImportError:
-    # 如果都找不到，手动定义一个空的 CPU 兼容类（防止报错）
-    import torch.nn as nn
-    class Text2SemanticDecoder(nn.Module):
-        def __init__(self, *args, **kwargs):
-            super().__init__()
-            print("⚠️ 使用了紧急备用 CPU Decoder 类")
-        def forward(self, *args, **kwargs):
-            print("⚠️ 紧急备用 Forward 被调用")
-            return None
-
-# 导出这个类，骗过调用者
-Text2SemanticDecoderFlashAttn = Text2SemanticDecoder
-"""
-
-# 递归搜索并覆盖
-target_filename = "t2s_model_flash_attn.py"
-hacked_count = 0
-
-for root, dirs, files in os.walk("."):
-    if target_filename in files:
-        full_path = os.path.join(root, target_filename)
-        print(f"🎯 锁定目标: {full_path}")
-        
-        try:
-            with open(full_path, "w", encoding="utf-8") as f:
-                f.write(CPU_HACK_CODE)
-            print("✅ 已执行物理覆盖 (写入 CPU 伪装代码)")
-            hacked_count += 1
-        except Exception as e:
-            print(f"❌ 覆盖失败: {e}")
-
-if hacked_count == 0:
-    print("⚠️ 警告：未找到任何 FlashAttn 文件！可能目录结构极其特殊。")
-else:
-    print(f"🎉 成功处理了 {hacked_count} 个 GPU 文件。")
-
-# ==========================================
-# 2. 基础环境设置
+# 1. 净化环境
 # ==========================================
 os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
 import torch
@@ -67,9 +12,50 @@ torch.cuda.device_count = lambda: 0
 def no_op(self, *args, **kwargs): return self
 torch.Tensor.cuda = no_op
 torch.nn.Module.cuda = no_op
+print("💉 CUDA 已屏蔽")
 
 # ==========================================
-# 3. 导入推理核心
+# 2. 物理层克隆：用 CPU 代码覆盖 GPU 代码
+# ==========================================
+print("🧬 启动代码克隆手术...")
+
+def find_file(name, search_path="."):
+    for root, dirs, files in os.walk(search_path):
+        if name in files:
+            return os.path.join(root, name)
+    return None
+
+# 1. 找到源文件 (CPU版)
+cpu_src = find_file("t2s_model.py")
+# 2. 找到目标文件 (GPU版)
+gpu_dst = find_file("t2s_model_flash_attn.py")
+
+if cpu_src and gpu_dst:
+    try:
+        # 读取 CPU 代码
+        with open(cpu_src, "r", encoding="utf-8") as f:
+            cpu_code = f.read()
+        
+        # 写入 GPU 文件
+        with open(gpu_dst, "w", encoding="utf-8") as f:
+            f.write(cpu_code)
+            
+        print(f"✅ 手术成功！\n源: {cpu_src}\n目标: {gpu_dst}")
+        print("现在系统加载 FlashAttn 模块时，实际上运行的是纯 CPU 代码。")
+        
+        # 🧹 清理 pycache 防止缓存作祟
+        cache_dir = os.path.join(os.path.dirname(gpu_dst), "__pycache__")
+        if os.path.exists(cache_dir):
+            shutil.rmtree(cache_dir)
+            print("🧹 已清理 __pycache__")
+            
+    except Exception as e:
+        print(f"❌ 文件操作失败: {e}")
+else:
+    print(f"⚠️ 未找到关键代码文件 (src:{cpu_src}, dst:{gpu_dst})")
+
+# ==========================================
+# 3. 导入核心逻辑
 # ==========================================
 sys.path.append(os.getcwd())
 
@@ -77,49 +63,51 @@ try:
     import inference_webui as core
     print("✅ 成功导入 inference_webui")
     
+    # 强制修改配置
     if hasattr(core, "is_half"): core.is_half = False
     if hasattr(core, "device"): core.device = "cpu"
         
-except ImportError:
-    print("❌ 找不到 inference_webui.py")
+except Exception as e:
+    print(f"❌ 导入失败: {e}")
+    # 打印一下当前目录，方便排查
+    print(f"当前目录文件: {os.listdir('.')}")
     sys.exit(1)
 
 # ==========================================
-# 4. 自动寻找模型文件
+# 4. 自动寻找模型
 # ==========================================
-def find_real_model(pattern, search_path="."):
-    candidates = []
-    for root, dirs, files in os.walk(search_path):
+def find_model_file(pattern):
+    for root, dirs, files in os.walk("."):
         for file in files:
             if pattern in file and not file.endswith(".lock") and not file.endswith(".metadata"):
                 path = os.path.join(root, file)
                 size_mb = os.path.getsize(path) / (1024 * 1024)
-                if size_mb > 10: 
-                    candidates.append((path, size_mb))
-    if candidates:
-        candidates.sort(key=lambda x: x[1], reverse=True)
-        return candidates[0][0]
+                if size_mb > 10: return path
     return None
 
-gpt_path = find_real_model("s1v3.ckpt") or find_real_model("s1bert")
-sovits_path = find_real_model("s2Gv2ProPlus.pth") or find_real_model("s2G")
+gpt_path = find_model_file("s1v3.ckpt") or find_model_file("s1bert")
+sovits_path = find_model_file("s2Gv2ProPlus.pth") or find_model_file("s2G")
 
 # ==========================================
 # 5. 加载模型
 # ==========================================
 try:
     if gpt_path and sovits_path:
-        core.is_half = False
-        if hasattr(core, "change_gpt_weights"): core.change_gpt_weights(gpt_path=gpt_path)
-        if hasattr(core, "change_sovits_weights"): core.change_sovits_weights(sovits_path=sovits_path)
-        print(f"🎉 模型加载成功！")
+        # 再次强制关闭半精度
+        core.is_half = False 
+        
+        if hasattr(core, "change_gpt_weights"):
+            core.change_gpt_weights(gpt_path=gpt_path)
+        if hasattr(core, "change_sovits_weights"):
+            core.change_sovits_weights(sovits_path=sovits_path)
+        print(f"🎉 模型加载成功！(CPU Mode)")
     else:
         print("❌ 未找到模型文件")
 except Exception as e:
     print(f"⚠️ 模型加载报错: {e}")
 
 # ==========================================
-# 6. 推理接口
+# 6. 推理逻辑
 # ==========================================
 import soundfile as sf
 import gradio as gr
@@ -127,7 +115,7 @@ import numpy as np
 
 REF_AUDIO = "ref.wav"
 REF_TEXT = "你好"
-REF_LANG = "中文"
+REF_LANG = "中文" 
 
 def run_predict(text):
     if not os.path.exists(REF_AUDIO): return None, "❌ 请上传 ref.wav"
@@ -152,6 +140,7 @@ def run_predict(text):
             sr, data = result_list[0]
             out_path = f"out_{os.urandom(4).hex()}.wav"
             sf.write(out_path, data, sr)
+            print(f"✅ 生成完毕: {out_path}")
             return out_path, "✅ 成功"
             
     except Exception as e:
@@ -163,13 +152,16 @@ def run_predict(text):
 # 7. 启动界面
 # ==========================================
 with gr.Blocks() as app:
-    gr.Markdown(f"### GPT-SoVITS V2 (CPU HACKED)")
+    gr.Markdown(f"### GPT-SoVITS V2 (CPU Clone Edition)")
+    
     with gr.Row():
-        inp = gr.Textbox(label="文本", value="这下舒服了，终于跑通了。")
+        inp = gr.Textbox(label="文本", value="这一波，稳了。")
         btn = gr.Button("生成")
+    
     with gr.Row():
         out = gr.Audio(label="结果")
         log = gr.Textbox(label="日志")
+        
     btn.click(run_predict, [inp], [out, log], api_name="predict")
 
 if __name__ == "__main__":