Update handler.py

handler.py

@@ -1,68 +1,29 @@
 import subprocess
 import sys
+import site
 
-# 🚀 强制在代码运行最开始，从 PyPI 热升级 transformers 和 accelerate 到最新版本
+# 🚀 1. 强制从 PyPI 热升级 transformers 和 accelerate 到最新版本
 try:
-    print("Starting hot-upgrade of transformers and accelerate...")
+    print("Starting hot-upgrade of transformers and accelerate...", flush=True)
     subprocess.run(
         [sys.executable, "-m", "pip", "install", "--upgrade", "transformers", "accelerate", "--user"], 
         check=True
     )
-    print("Hot-upgrade completed successfully!")
+    print("Hot-upgrade completed successfully!", flush=True)
 except Exception as e:
-    print(f"Warning during hot-upgrade: {e}")
+    print(f"Warning during hot-upgrade: {e}", flush=True)
 
-# ----------------- 升级完成后再导入库 -----------------
-import torch
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-class EndpointHandler:
-    def __init__(self, path=""):
-        # 1. 使用 trust_remote_code=True 加载 Tokenizer 和 Model
-        self.tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)
-        self.model = AutoModelForCausalLM.from_pretrained(
-            path,
-            torch_dtype=torch.bfloat16,
-            device_map="auto",
-            trust_remote_code=True
-        )
+# 🚀 2. 强制将 user site-packages 插入到 sys.path 的最前面，覆盖系统自带的旧版库
+user_site = site.getusersitepackages()
+if user_site in sys.path:
+    sys.path.remove(user_site)
+sys.path.insert(0, user_site)
 
-    def __call__(self, data):
-        # 2. 获取输入和参数
-        inputs = data.get("inputs", "")
-        messages = data.get("messages", None)
-        parameters = data.get("parameters", {})
-        
-        # 3. 兼容 OpenAI 格式的 Chat completion
-        if messages is not None:
-            formatted_input = self.tokenizer.apply_chat_template(
-                messages, tokenize=False, add_generation_prompt=True
-            )
-        else:
-            formatted_input = inputs
+# 🚀 3. 清理内存中已加载的旧版模块，强制 Python 重新从新路径导入
+for mod in list(sys.modules.keys()):
+    if mod.startswith("transformers") or mod.startswith("accelerate"):
+        del sys.modules[mod]
 
-        # 4. 编码输入
-        inputs_tokenized = self.tokenizer([formatted_input], return_tensors="pt").to("cuda")
-        
-        # 5. 解析参数
-        max_new_tokens = parameters.get("max_new_tokens", parameters.get("max_tokens", 128))
-        temperature = parameters.get("temperature", 0.7)
-        top_p = parameters.get("top_p", 0.9)
-        do_sample = parameters.get("do_sample", True)
-
-        # 6. 进行推理
-        with torch.no_grad():
-            outputs = self.model.generate(
-                **inputs_tokenized,
-                max_new_tokens=max_new_tokens,
-                temperature=temperature,
-                top_p=top_p,
-                do_sample=do_sample
-            )
-        
-        # 7. 只解码新生成的回复部分
-        input_length = inputs_tokenized.input_ids.shape[1]
-        reply = self.tokenizer.decode(outputs[0][input_length:], skip_special_tokens=True)
-        
-        # 返回符合 Hugging Face 规范的格式
-        return [{"generated_text": reply}]
+# ----------------- 路径重排和模块清理完成后，导入新版库 -----------------
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer