Update handler.py
Browse files- handler.py +80 -68
handler.py
CHANGED
@@ -9,54 +9,92 @@ logger = logging.getLogger(__name__)
|
|
9 |
class EndpointHandler:
|
10 |
def __init__(self, model_dir: str = None):
|
11 |
self.model_dir = model_dir if model_dir else "homer7676/FrierenChatbotV1"
|
12 |
-
self.tokenizer = None
|
13 |
-
self.model = None
|
14 |
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
15 |
logger.info(f"初始化 EndpointHandler,使用設備: {self.device}")
|
16 |
-
|
17 |
-
|
18 |
-
try:
|
19 |
-
inputs = self.preprocess(data)
|
20 |
-
outputs = self.inference(inputs)
|
21 |
-
# 確保輸出不為空
|
22 |
-
if not outputs or "generated_text" not in outputs:
|
23 |
-
raise ValueError("No text was generated")
|
24 |
-
return [outputs]
|
25 |
-
except Exception as e:
|
26 |
-
logger.error(f"處理過程錯誤: {str(e)}")
|
27 |
-
return [{"error": str(e)}]
|
28 |
-
|
29 |
-
def initialize(self, context):
|
30 |
-
logger.info("開始初始化模型")
|
31 |
try:
|
|
|
32 |
self.tokenizer = AutoTokenizer.from_pretrained(
|
33 |
self.model_dir,
|
34 |
trust_remote_code=True
|
35 |
)
|
36 |
-
|
37 |
if self.tokenizer.pad_token is None:
|
38 |
self.tokenizer.pad_token = self.tokenizer.eos_token
|
39 |
-
|
40 |
self.model = AutoModelForCausalLM.from_pretrained(
|
41 |
self.model_dir,
|
42 |
trust_remote_code=True,
|
43 |
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
|
44 |
).to(self.device)
|
45 |
-
|
46 |
self.model.eval()
|
47 |
-
logger.info("
|
|
|
48 |
except Exception as e:
|
49 |
-
logger.error(f"
|
50 |
raise
|
51 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
def inference(self, inputs: Dict[str, Any]) -> Dict[str, str]:
|
53 |
logger.info("開始執行推理")
|
54 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
message = inputs.get("message", "")
|
56 |
context = inputs.get("context", "")
|
57 |
-
logger.info(f"
|
58 |
-
|
59 |
-
# 構建提示詞
|
60 |
prompt = f"""你是芙莉蓮,需要遵守以下規則回答:
|
61 |
1. 身份設定:
|
62 |
- 千年精靈魔法師
|
@@ -75,69 +113,43 @@ class EndpointHandler:
|
|
75 |
用戶:{message}
|
76 |
芙莉蓮:"""
|
77 |
|
78 |
-
#
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
prompt,
|
84 |
-
add_special_tokens=True,
|
85 |
return_tensors="pt",
|
86 |
padding=True,
|
87 |
truncation=True,
|
88 |
max_length=2048
|
89 |
-
)
|
90 |
-
|
91 |
-
# 移動到正確的設備
|
92 |
-
input_ids = encoding["input_ids"].to(self.device)
|
93 |
-
attention_mask = encoding["attention_mask"].to(self.device)
|
94 |
-
|
95 |
-
logger.info(f"輸入 token 數量: {input_ids.shape[-1]}")
|
96 |
|
97 |
-
# 生成回應
|
98 |
with torch.no_grad():
|
99 |
outputs = self.model.generate(
|
100 |
-
|
101 |
-
|
102 |
-
max_new_tokens=256,
|
103 |
temperature=0.7,
|
104 |
top_p=0.9,
|
105 |
top_k=50,
|
106 |
do_sample=True,
|
107 |
pad_token_id=self.tokenizer.pad_token_id,
|
108 |
-
eos_token_id=self.tokenizer.eos_token_id
|
109 |
-
num_return_sequences=1,
|
110 |
-
no_repeat_ngram_size=3
|
111 |
)
|
112 |
-
|
113 |
-
logger.info(f"生成的 token 數量: {outputs.shape[-1]}")
|
114 |
|
115 |
-
|
116 |
-
|
117 |
|
118 |
-
# 分離出模型的回應部分
|
119 |
-
if "芙莉蓮:" in full_response:
|
120 |
-
response = full_response.split("芙莉蓮:")[-1].strip()
|
121 |
-
else:
|
122 |
-
response = full_response.split("用戶:")[-1].strip()
|
123 |
-
|
124 |
-
logger.info(f"生成回應長度: {len(response)}")
|
125 |
-
|
126 |
-
# 確保回應不為空
|
127 |
if not response:
|
128 |
-
response = "
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
}
|
133 |
|
134 |
except Exception as e:
|
135 |
logger.error(f"推理過程錯誤: {str(e)}")
|
136 |
-
|
137 |
|
138 |
def preprocess(self, data: Dict[str, Any]) -> Dict[str, Any]:
|
139 |
logger.info(f"預處理輸入數據: {data}")
|
140 |
-
|
141 |
-
if not isinstance(inputs, dict):
|
142 |
-
inputs = {"message": inputs}
|
143 |
-
return inputs
|
|
|
9 |
class EndpointHandler:
|
10 |
def __init__(self, model_dir: str = None):
|
11 |
self.model_dir = model_dir if model_dir else "homer7676/FrierenChatbotV1"
|
|
|
|
|
12 |
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
13 |
logger.info(f"初始化 EndpointHandler,使用設備: {self.device}")
|
14 |
+
|
15 |
+
# 在初始化時就載入模型和 tokenizer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
try:
|
17 |
+
logger.info("開始載入 tokenizer 和模型")
|
18 |
self.tokenizer = AutoTokenizer.from_pretrained(
|
19 |
self.model_dir,
|
20 |
trust_remote_code=True
|
21 |
)
|
|
|
22 |
if self.tokenizer.pad_token is None:
|
23 |
self.tokenizer.pad_token = self.tokenizer.eos_token
|
24 |
+
|
25 |
self.model = AutoModelForCausalLM.from_pretrained(
|
26 |
self.model_dir,
|
27 |
trust_remote_code=True,
|
28 |
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
|
29 |
).to(self.device)
|
|
|
30 |
self.model.eval()
|
31 |
+
logger.info("模型和 tokenizer 載入完成")
|
32 |
+
|
33 |
except Exception as e:
|
34 |
+
logger.error(f"初始化錯誤: {str(e)}")
|
35 |
raise
|
36 |
|
37 |
+
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, str]]:
|
38 |
+
try:
|
39 |
+
# 確保 tokenizer 和 model 已經初始化
|
40 |
+
if self.tokenizer is None or self.model is None:
|
41 |
+
raise RuntimeError("Tokenizer or model not initialized")
|
42 |
+
|
43 |
+
inputs = self.preprocess(data)
|
44 |
+
outputs = self.inference(inputs)
|
45 |
+
return [outputs]
|
46 |
+
except Exception as e:
|
47 |
+
logger.error(f"處理過程錯誤: {str(e)}")
|
48 |
+
return [{"error": str(e)}]
|
49 |
+
|
50 |
+
def initialize(self, context):
|
51 |
+
"""確保模型已初始化"""
|
52 |
+
if self.tokenizer is None or self.model is None:
|
53 |
+
logger.info("在 initialize 中重新初始化模型")
|
54 |
+
try:
|
55 |
+
self.tokenizer = AutoTokenizer.from_pretrained(
|
56 |
+
self.model_dir,
|
57 |
+
trust_remote_code=True
|
58 |
+
)
|
59 |
+
if self.tokenizer.pad_token is None:
|
60 |
+
self.tokenizer.pad_token = self.tokenizer.eos_token
|
61 |
+
|
62 |
+
self.model = AutoModelForCausalLM.from_pretrained(
|
63 |
+
self.model_dir,
|
64 |
+
trust_remote_code=True,
|
65 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
|
66 |
+
).to(self.device)
|
67 |
+
self.model.eval()
|
68 |
+
logger.info("模型重新初始化完成")
|
69 |
+
except Exception as e:
|
70 |
+
logger.error(f"模型重新初始化錯誤: {str(e)}")
|
71 |
+
raise
|
72 |
+
|
73 |
def inference(self, inputs: Dict[str, Any]) -> Dict[str, str]:
|
74 |
logger.info("開始執行推理")
|
75 |
try:
|
76 |
+
# 檢查輸入格式
|
77 |
+
if isinstance(inputs, str):
|
78 |
+
try:
|
79 |
+
import json
|
80 |
+
inputs = json.loads(inputs)
|
81 |
+
except json.JSONDecodeError:
|
82 |
+
inputs = {"message": inputs}
|
83 |
+
|
84 |
+
# 提取消息和上下文
|
85 |
+
if isinstance(inputs, dict) and "inputs" in inputs:
|
86 |
+
inputs = inputs["inputs"]
|
87 |
+
if isinstance(inputs, str):
|
88 |
+
try:
|
89 |
+
import json
|
90 |
+
inputs = json.loads(inputs)
|
91 |
+
except json.JSONDecodeError:
|
92 |
+
inputs = {"message": inputs}
|
93 |
+
|
94 |
message = inputs.get("message", "")
|
95 |
context = inputs.get("context", "")
|
96 |
+
logger.info(f"處理消息: {message}, 上下文: {context}")
|
97 |
+
|
|
|
98 |
prompt = f"""你是芙莉蓮,需要遵守以下規則回答:
|
99 |
1. 身份設定:
|
100 |
- 千年精靈魔法師
|
|
|
113 |
用戶:{message}
|
114 |
芙莉蓮:"""
|
115 |
|
116 |
+
# 確保 tokenizer 存在
|
117 |
+
if self.tokenizer is None:
|
118 |
+
raise RuntimeError("Tokenizer not initialized")
|
119 |
+
|
120 |
+
tokens = self.tokenizer(
|
121 |
prompt,
|
|
|
122 |
return_tensors="pt",
|
123 |
padding=True,
|
124 |
truncation=True,
|
125 |
max_length=2048
|
126 |
+
).to(self.device)
|
|
|
|
|
|
|
|
|
|
|
|
|
127 |
|
|
|
128 |
with torch.no_grad():
|
129 |
outputs = self.model.generate(
|
130 |
+
**tokens,
|
131 |
+
max_new_tokens=150,
|
|
|
132 |
temperature=0.7,
|
133 |
top_p=0.9,
|
134 |
top_k=50,
|
135 |
do_sample=True,
|
136 |
pad_token_id=self.tokenizer.pad_token_id,
|
137 |
+
eos_token_id=self.tokenizer.eos_token_id
|
|
|
|
|
138 |
)
|
|
|
|
|
139 |
|
140 |
+
response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
|
141 |
+
response = response.split("芙莉蓮:")[-1].strip()
|
142 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
143 |
if not response:
|
144 |
+
response = "唔...讓我思考一下如何回答你的問題。"
|
145 |
+
|
146 |
+
logger.info(f"生成回應: {response}")
|
147 |
+
return {"generated_text": response}
|
|
|
148 |
|
149 |
except Exception as e:
|
150 |
logger.error(f"推理過程錯誤: {str(e)}")
|
151 |
+
return {"error": str(e)}
|
152 |
|
153 |
def preprocess(self, data: Dict[str, Any]) -> Dict[str, Any]:
|
154 |
logger.info(f"預處理輸入數據: {data}")
|
155 |
+
return data
|
|
|
|
|
|