Update app.py

app.py

@@ -1,9 +1,23 @@
 import os
 import sys
-import logging
+import types
 
-# --- 1. 基础环境设置 ---
-os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
+# ==========================================
+# 1. 核心欺骗：制造“假”的 Flash Attention
+# ==========================================
+# 这一步必须在所有 imports 之前！
+# 我们创建一个空的模块，骗过系统的 import 检查
+# 但因为里面没有 functional 接口，模型会报错并回退到普通模式
+dummy_module = types.ModuleType("flash_attn")
+sys.modules["flash_attn"] = dummy_module
+sys.modules["flash_attn.flash_attn_interface"] = dummy_module
+
+print("💉 已注入 Flash Attention 假模块，强制开启 CPU 兼容模式。")
+
+# ==========================================
+# 2. 屏蔽 CUDA
+# ==========================================
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
 import torch
 torch.cuda.is_available = lambda: False
 torch.cuda.device_count = lambda: 0
@@ -11,29 +25,29 @@ def no_op(self, *args, **kwargs): return self
 torch.Tensor.cuda = no_op
 torch.nn.Module.cuda = no_op
 
-print("💉 CUDA 补丁已注入")
+print("💉 CUDA 已屏蔽。")
 
-# --- 2. 导入核心逻辑 ---
+# ==========================================
+# 3. 导入核心逻辑
+# ==========================================
 sys.path.append(os.getcwd())
 
 try:
     import inference_webui as core
     print("✅ 成功导入 inference_webui")
     
-    # 🛑 关键修正：强制关闭半精度 (CPU 不支持 FP16，也能顺便避开 Flash Attention)
+    # 强制关闭半精度
     if hasattr(core, "is_half"):
         core.is_half = False
-        print("✅ 强制禁用半精度 (is_half = False)")
-    
-    if hasattr(core, "device"):
-        core.device = "cpu"
-        print("✅ 强制指定设备 (device = cpu)")
+        print("✅ 强制禁用半精度")
         
 except ImportError:
-    print("❌ 严重错误：找不到 inference_webui.py")
+    print("❌ 找不到 inference_webui.py")
     sys.exit(1)
 
-# --- 3. 自动寻找模型 ---
+# ==========================================
+# 4. 自动寻找模型
+# ==========================================
 def find_real_model(pattern, search_path="."):
     candidates = []
     for root, dirs, files in os.walk(search_path):
@@ -51,31 +65,34 @@ def find_real_model(pattern, search_path="."):
 gpt_path = find_real_model("s1v3.ckpt") or find_real_model("s1bert")
 sovits_path = find_real_model("s2Gv2ProPlus.pth") or find_real_model("s2G")
 
-# --- 4. 加载模型 ---
+# ==========================================
+# 5. 加载模型
+# ==========================================
 try:
     if gpt_path and sovits_path:
-        # 再次确保加载模型时不会启用半精度
+        # 再次确保
         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！")
+        print(f"🎉 模型加载成功！(CPU模式)")
     else:
         print("❌ 未找到模型文件")
 except Exception as e:
     print(f"⚠️ 模型加载报错: {e}")
 
-# --- 5. 推理逻辑 ---
+# ==========================================
+# 6. 推理逻辑
+# ==========================================
 import soundfile as sf
 import gradio as gr
 import numpy as np
 
 REF_AUDIO = "ref.wav"
 REF_TEXT = "你好"
-# 🛑 关键修正：语言必须是中文
-REF_LANG = "中文" 
+REF_LANG = "中文" # 必须是中文
 
 def run_predict(text):
     if not os.path.exists(REF_AUDIO):
@@ -83,18 +100,16 @@ def run_predict(text):
     
     print(f"📥 任务: {text}")
     try:
-        # 自动识别函数
         inference_func = getattr(core, "get_tts_model", getattr(core, "get_tts_wav", None))
-        if not inference_func:
-            return None, "❌ 找不到推理函数"
+        if not inference_func: return None, "❌ 找不到推理函数"
 
         # 核心调用
         generator = inference_func(
             ref_wav_path=REF_AUDIO,
             prompt_text=REF_TEXT,
-            prompt_language=REF_LANG, # 中文
+            prompt_language=REF_LANG,
             text=text,
-            text_language="中文",     # 中文
+            text_language="中文",
             how_to_cut="凑四句一切",
             top_k=5, top_p=1, temperature=1, ref_free=False
         )
@@ -112,12 +127,15 @@ def run_predict(text):
         traceback.print_exc()
         return None, f"💥 报错: {e}"
 
-# --- 6. 界面 ---
+# ==========================================
+# 7. 界面
+# ==========================================
 with gr.Blocks() as app:
-    gr.Markdown(f"### GPT-SoVITS V2 (Final CPU)")
+    gr.Markdown(f"### GPT-SoVITS CPU 终极版")
+    gr.Markdown(f"Status: FlashAttn Disabled, CUDA Disabled, FP32 Mode")
     
     with gr.Row():
-        inp = gr.Textbox(label="文本", value="这次一定行，不行我就吃键盘。")
+        inp = gr.Textbox(label="文本", value="如果听到这句话，说明你成功了！")
         btn = gr.Button("生成")
     
     with gr.Row():