返回提交历史
Modified
g4f/Provider/Qwen.py
+68
-59
Modified
g4f/Provider/__init__.py
+1
-0
Modified
g4f/Provider/needs_auth/Nvidia.py
+3
-8
Modified
g4f/providers/response.py
+6
-0
XFEstudio/gpt4free
feat: add Qwen provider with conversation support and stream handling
- Added `Qwen` to `g4f/Provider/__init__.py` for provider registration - Created new Qwen provider in `g4f/Provider/Qwen.py` using `AsyncGeneratorProvider` - Implemented conversation state via new `JsonConversation` argument - Replaced raw `print` statements with `debug.log` for internal logging - Introduced `get_last_user_message()` for improved prompt extraction - Added support for `Reasoning` and `Usage` response types during SSE parsing - Replaced manual SSE parsing with `sse_stream()` utility from `requests` - Added `active_by_default = True` to `Qwen` and modified related headers - Tracked message and parent IDs for contextual threading - Updated `Usage` class in `g4f/providers/response.py` to support `input_tokens` and `output_tokens` - Refactored Nvidia provider: removed unused attributes and set `models_needs_auth = True
31fee02c
代码差异
4 个文件
+78
-67
@@ -1,3 +1,5 @@
1
from __future__ import annotations
2
1
3
import asyncio
2
4
import json
3
5
import re
@@ -6,8 +8,12 @@ from time import time
6
8
7
9
import aiohttp
8
10
from ..errors import RateLimitError
9
from .base_provider import AsyncGeneratorProvider, ProviderModelMixin
10
11
from ..typing import AsyncResult, Messages
12
from ..providers.response import JsonConversation, Reasoning, Usage
13
from ..requests import sse_stream
14
from .base_provider import AsyncGeneratorProvider, ProviderModelMixin
15
from .helper import get_last_user_message
16
from .. import debug
11
17
12
18
class Qwen(AsyncGeneratorProvider, ProviderModelMixin):
13
19
"""
@@ -16,6 +22,7 @@ class Qwen(AsyncGeneratorProvider, ProviderModelMixin):
16
22
"""
17
23
url = "https://chat.qwen.ai"
18
24
working = True
25
active_by_default = True
19
26
supports_stream = True
20
27
supports_message_history = False
21
28
@@ -46,6 +53,7 @@ class Qwen(AsyncGeneratorProvider, ProviderModelMixin):
46
53
cls,
47
54
model: str,
48
55
messages: Messages,
56
conversation: JsonConversation = None,
49
57
proxy: str = None,
50
58
timeout: int = 120,
51
59
stream: bool = True,
@@ -70,13 +78,13 @@ class Qwen(AsyncGeneratorProvider, ProviderModelMixin):
70
78
'Source': 'web'
71
79
}
72
80
73
prompt = messages[-1]["content"]
81
prompt = get_last_user_message(messages)
74
82
75
83
async with aiohttp.ClientSession(headers=headers) as session:
76
84
for attempt in range(5):
77
85
try:
78
86
if not cls._midtoken:
79
print("[Qwen] INFO: No active midtoken. Fetching a new one...")
87
debug.log("[Qwen] INFO: No active midtoken. Fetching a new one...")
80
88
async with session.get('https://sg-wum.alibaba.com/w/wu.json', proxy=proxy) as r:
81
89
r.raise_for_status()
82
90
text = await r.text()
@@ -85,42 +93,48 @@ class Qwen(AsyncGeneratorProvider, ProviderModelMixin):
85
93
raise RuntimeError("Failed to extract bx-umidtoken.")
86
94
cls._midtoken = match.group(1)
87
95
cls._midtoken_uses = 1
88
print(f"[Qwen] INFO: New midtoken obtained. Use count: {cls._midtoken_uses}. Midtoken: {cls._midtoken}")
96
debug.log(f"[Qwen] INFO: New midtoken obtained. Use count: {cls._midtoken_uses}. Midtoken: {cls._midtoken}")
89
97
else:
90
98
cls._midtoken_uses += 1
91
print(f"[Qwen] INFO: Reusing midtoken. Use count: {cls._midtoken_uses}")
99
debug.log(f"[Qwen] INFO: Reusing midtoken. Use count: {cls._midtoken_uses}")
92
100
93
101
req_headers = session.headers.copy()
94
102
req_headers['bx-umidtoken'] = cls._midtoken
95
103
req_headers['bx-v'] = '2.5.31'
96
97
chat_payload = {
98
"title": "New Chat",
99
"models": [model_name],
100
"chat_mode": "normal",
101
"chat_type": "t2t",
102
"timestamp": int(time() * 1000)
103
}
104
async with session.post(
105
f'{cls.url}/api/v2/chats/new', json=chat_payload, headers=req_headers, proxy=proxy
106
) as resp:
107
resp.raise_for_status()
108
data = await resp.json()
109
if not (data.get('success') and data['data'].get('id')):
110
raise RuntimeError(f"Failed to create chat: {data}")
111
chat_id = data['data']['id']
112
104
message_id = str(uuid.uuid4())
105
parent_id = None
106
if conversation is None:
107
chat_payload = {
108
"title": "New Chat",
109
"models": [model_name],
110
"chat_mode": "normal",
111
"chat_type": "t2t",
112
"timestamp": int(time() * 1000)
113
}
114
async with session.post(
115
f'{cls.url}/api/v2/chats/new', json=chat_payload, headers=req_headers, proxy=proxy
116
) as resp:
117
resp.raise_for_status()
118
data = await resp.json()
119
if not (data.get('success') and data['data'].get('id')):
120
raise RuntimeError(f"Failed to create chat: {data}")
121
conversation = JsonConversation(
122
chat_id=data['data']['id'],
123
cookies={key: value for key, value in resp.cookies.items()},
124
parent_id=None
125
)
126
113
127
msg_payload = {
114
128
"stream": stream,
115
129
"incremental_output": stream,
116
"chat_id": chat_id,
130
"chat_id": conversation.chat_id,
117
131
"chat_mode": "normal",
118
132
"model": model_name,
119
"parent_id": None,
133
"parent_id": conversation.parent_id,
120
134
"messages": [
121
135
{
122
"fid": str(uuid.uuid4()),
123
"parentId": None,
136
"fid": message_id,
137
"parentId": conversation.parent_id,
124
138
"childrenIds": [],
125
139
"role": "user",
126
140
"content": prompt,
@@ -145,50 +159,45 @@ class Qwen(AsyncGeneratorProvider, ProviderModelMixin):
145
159
}
146
160
147
161
async with session.post(
148
f'{cls.url}/api/v2/chat/completions?chat_id={chat_id}', json=msg_payload,
149
headers=req_headers, proxy=proxy, timeout=timeout
162
f'{cls.url}/api/v2/chat/completions?chat_id={conversation.chat_id}', json=msg_payload,
163
headers=req_headers, proxy=proxy, timeout=timeout, cookies=conversation.cookies
150
164
) as resp:
151
165
first_line = await resp.content.readline()
152
166
line_str = first_line.decode().strip()
153
167
if line_str.startswith('{'):
154
error_data = json.loads(line_str)
155
if error_data.get("data", {}).get("code") == "RateLimited":
156
raise RuntimeError("RateLimited by JSON response")
157
158
buffer = first_line
168
data = json.loads(line_str)
169
if data.get("data", {}).get("code"):
170
raise RuntimeError(f"Response: {data}")
171
conversation.parent_id = data.get("response.created", {}).get("response_id")
172
yield conversation
173
159
174
thinking_started = False
160
async for chunk in resp.content:
161
buffer += chunk
162
while b'\n' in buffer:
163
line, buffer = buffer.split(b'\n', 1)
164
line_str = line.decode().strip()
165
if not line_str.startswith("data: "): continue
166
try:
167
data_json = json.loads(line_str.lstrip("data: "))
168
choices = data_json.get("choices", [])
169
if not choices: continue
170
delta = choices[0].get("delta", {})
171
phase = delta.get("phase")
172
content = delta.get("content")
173
if phase == "think" and not thinking_started:
174
yield "<think>"
175
thinking_started = True
176
elif phase == "answer" and thinking_started:
177
yield "</think>"
178
thinking_started = False
179
if content:
180
yield content
181
except (json.JSONDecodeError, KeyError, IndexError):
182
continue
183
if thinking_started:
184
yield "</think>"
175
usage = None
176
async for chunk in sse_stream(resp):
177
try:
178
usage = chunk.get("usage", usage)
179
choices = chunk.get("choices", [])
180
if not choices: continue
181
delta = choices[0].get("delta", {})
182
phase = delta.get("phase")
183
content = delta.get("content")
184
if phase == "think" and not thinking_started:
185
thinking_started = True
186
elif phase == "answer" and thinking_started:
187
thinking_started = False
188
if content:
189
yield Reasoning(content) if thinking_started else content
190
except (json.JSONDecodeError, KeyError, IndexError):
191
continue
192
if usage:
193
yield Usage(**usage)
185
194
return
186
195
187
196
except (aiohttp.ClientResponseError, RuntimeError) as e:
188
197
is_rate_limit = (isinstance(e, aiohttp.ClientResponseError) and e.status == 429) or \
189
198
("RateLimited" in str(e))
190
199
if is_rate_limit:
191
print(f"[Qwen] WARNING: Rate limit detected (attempt {attempt + 1}/5). Invalidating current midtoken.")
200
debug.log(f"[Qwen] WARNING: Rate limit detected (attempt {attempt + 1}/5). Invalidating current midtoken.")
192
201
cls._midtoken = None
193
202
cls._midtoken_uses = 0
194
203
await asyncio.sleep(2)
@@ -54,6 +54,7 @@ from .PerplexityLabs import PerplexityLabs
54
54
from .PollinationsAI import PollinationsAI
55
55
from .PollinationsImage import PollinationsImage
56
56
from .Startnest import Startnest
57
from .Qwen import Qwen
57
58
from .TeachAnything import TeachAnything
58
59
from .WeWordle import WeWordle
59
60
from .YouTube import YouTube
@@ -9,12 +9,7 @@ class Nvidia(OpenaiTemplate):
9
9
login_url = "https://google.com"
10
10
url = "https://build.nvidia.com"
11
11
working = True
12
active_by_default = True
12
13
needs_auth = True
13
supports_stream = True
14
supports_system_message = True
15
supports_message_history = True
16
default_model = DEFAULT_MODEL.split("/")[-1]
17
18
@classmethod
19
def get_model(cls, model: str, **kwargs) -> str:
20
return super().get_model(model, **kwargs)
14
models_needs_auth = True
15
default_model = DEFAULT_MODEL
@@ -170,12 +170,18 @@ class Usage(JsonMixin, HiddenResponse):
170
170
self,
171
171
promptTokens: int = None,
172
172
completionTokens: int = None,
173
input_tokens: int = None,
174
output_tokens: int = None,
173
175
**kwargs
174
176
):
175
177
if promptTokens is not None:
176
178
kwargs["prompt_tokens"] = promptTokens
177
179
if completionTokens is not None:
178
180
kwargs["completion_tokens"] = completionTokens
181
if input_tokens is not None:
182
kwargs["prompt_tokens"] = input_tokens
183
if output_tokens is not None:
184
kwargs["completion_tokens"] = output_tokens
179
185
if "total_tokens" not in kwargs and "prompt_tokens" in kwargs and "completion_tokens" in kwargs:
180
186
kwargs["total_tokens"] = kwargs["prompt_tokens"] + kwargs["completion_tokens"]
181
187
return super().__init__(**kwargs)