Update app.py
Browse files
app.py
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
import os
|
2 |
import json
|
3 |
import uuid
|
|
|
4 |
from datetime import datetime
|
5 |
from flask import Flask, request, Response, jsonify
|
6 |
import socketio
|
@@ -76,10 +77,11 @@ def normalize_content(content):
|
|
76 |
|
77 |
def calculate_tokens(text):
|
78 |
"""
|
79 |
-
|
80 |
-
这里假设平均每 4 个字符代表 1 个 token。
|
81 |
"""
|
82 |
-
|
|
|
|
|
83 |
|
84 |
@app.route('/')
|
85 |
def root():
|
@@ -122,7 +124,7 @@ def messages():
|
|
122 |
|
123 |
msg_id = str(uuid.uuid4())
|
124 |
response_event = Event()
|
125 |
-
response_text = []
|
126 |
|
127 |
if not stream:
|
128 |
# 处理 stream 为 false 的情况
|
@@ -142,7 +144,7 @@ def messages():
|
|
142 |
"model": model, # 动态模型
|
143 |
"stop_reason": None,
|
144 |
"stop_sequence": None,
|
145 |
-
"usage": {"input_tokens": input_tokens, "output_tokens": 0}, # 动态 input_tokens
|
146 |
},
|
147 |
})
|
148 |
yield create_event("content_block_start", {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}})
|
@@ -220,20 +222,15 @@ def messages():
|
|
220 |
if sio.connected:
|
221 |
sio.disconnect()
|
222 |
|
223 |
-
#
|
224 |
-
|
225 |
-
# 动态计算 output_tokens
|
226 |
-
output_tokens = calculate_tokens(output_text)
|
227 |
|
228 |
-
# 生成 content_block_stop 事件
|
229 |
yield create_event("content_block_stop", {"type": "content_block_stop", "index": 0})
|
230 |
-
# 生成 message_delta 事件,包含 output_tokens
|
231 |
yield create_event("message_delta", {
|
232 |
"type": "message_delta",
|
233 |
"delta": {"stop_reason": "end_turn", "stop_sequence": None},
|
234 |
-
"usage": {"output_tokens": output_tokens}, # 动态 output_tokens
|
235 |
})
|
236 |
-
# 生成 message_stop 事件
|
237 |
yield create_event("message_stop", {"type": "message_stop"}) # 确保发送 message_stop 事件
|
238 |
|
239 |
return Response(generate(), content_type='text/event-stream')
|
@@ -300,14 +297,12 @@ def handle_non_stream(previous_messages, msg_id, model, input_tokens):
|
|
300 |
# 等待响应完成
|
301 |
response_event.wait(timeout=30)
|
302 |
|
303 |
-
#
|
304 |
-
|
305 |
-
# 动态计算 output_tokens
|
306 |
-
output_tokens = calculate_tokens(output_text)
|
307 |
|
308 |
# 生成完整的响应
|
309 |
full_response = {
|
310 |
-
"content": [{"text":
|
311 |
"id": msg_id,
|
312 |
"model": model, # 动态模型
|
313 |
"role": "assistant",
|
|
|
1 |
import os
|
2 |
import json
|
3 |
import uuid
|
4 |
+
import re
|
5 |
from datetime import datetime
|
6 |
from flask import Flask, request, Response, jsonify
|
7 |
import socketio
|
|
|
77 |
|
78 |
def calculate_tokens(text):
|
79 |
"""
|
80 |
+
计算输入文本的 token 数量。我们通过正则表达式来进行 tokenization,确保标点符号和单词都被正确计算为 token。
|
|
|
81 |
"""
|
82 |
+
# 使用正则表达式匹配单词和标点符号
|
83 |
+
tokens = re.findall(r"\w+|[^\w\s]", text)
|
84 |
+
return len(tokens)
|
85 |
|
86 |
@app.route('/')
|
87 |
def root():
|
|
|
124 |
|
125 |
msg_id = str(uuid.uuid4())
|
126 |
response_event = Event()
|
127 |
+
response_text = []
|
128 |
|
129 |
if not stream:
|
130 |
# 处理 stream 为 false 的情况
|
|
|
144 |
"model": model, # 动态模型
|
145 |
"stop_reason": None,
|
146 |
"stop_sequence": None,
|
147 |
+
"usage": {"input_tokens": input_tokens, "output_tokens": 0}, # 动态 input_tokens,output_tokens 初始化为 0
|
148 |
},
|
149 |
})
|
150 |
yield create_event("content_block_start", {"type": "content_block_start", "index": 0, "content_block": {"type": "text", "text": ""}})
|
|
|
222 |
if sio.connected:
|
223 |
sio.disconnect()
|
224 |
|
225 |
+
# 计算输出 token 数量
|
226 |
+
output_tokens = calculate_tokens(''.join(response_text))
|
|
|
|
|
227 |
|
|
|
228 |
yield create_event("content_block_stop", {"type": "content_block_stop", "index": 0})
|
|
|
229 |
yield create_event("message_delta", {
|
230 |
"type": "message_delta",
|
231 |
"delta": {"stop_reason": "end_turn", "stop_sequence": None},
|
232 |
+
"usage": {"input_tokens": input_tokens, "output_tokens": output_tokens}, # 动态 output_tokens
|
233 |
})
|
|
|
234 |
yield create_event("message_stop", {"type": "message_stop"}) # 确保发送 message_stop 事件
|
235 |
|
236 |
return Response(generate(), content_type='text/event-stream')
|
|
|
297 |
# 等待响应完成
|
298 |
response_event.wait(timeout=30)
|
299 |
|
300 |
+
# 计算输出 token 数量
|
301 |
+
output_tokens = calculate_tokens(''.join(response_text))
|
|
|
|
|
302 |
|
303 |
# 生成完整的响应
|
304 |
full_response = {
|
305 |
+
"content": [{"text": ''.join(response_text), "type": "text"}], # 合并所有文本块
|
306 |
"id": msg_id,
|
307 |
"model": model, # 动态模型
|
308 |
"role": "assistant",
|