fb700 commited on
Commit
f04af32
1 Parent(s): 99cd7d3

Upload 7 files

Browse files
Files changed (3) hide show
  1. README.md +7 -6
  2. app bak.py +386 -0
  3. app.py +76 -56
README.md CHANGED
@@ -1,12 +1,13 @@
1
  ---
2
- title: chatglm2 6b int4
3
- emoji: 🌖
4
- colorFrom: purple
5
- colorTo: blue
6
  sdk: gradio
7
- sdk_version: 3.35.2
8
  app_file: app.py
9
- pinned: true
 
10
  ---
11
 
12
  Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
1
  ---
2
+ title: Chatglm Fitness RLHF
3
+ emoji: 🐠
4
+ colorFrom: yellow
5
+ colorTo: yellow
6
  sdk: gradio
7
+ sdk_version: 3.38.0
8
  app_file: app.py
9
+ pinned: false
10
+ license: apache-2.0
11
  ---
12
 
13
  Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
app bak.py ADDED
@@ -0,0 +1,386 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Credit to https://github.com/THUDM/ChatGLM2-6B/blob/main/web_demo.py while mistakes are mine."""
2
+ # pylint: disable=broad-exception-caught, redefined-outer-name, missing-function-docstring, missing-module-docstring, too-many-arguments, line-too-long, invalid-name, redefined-builtin, redefined-argument-from-local
3
+ # import gradio as gr
4
+
5
+ # model_name = "fb700/chatglm-fitness-RLHF"
6
+ # gr.load(model_name).lauch()
7
+
8
+ # %%writefile demo-4bit.py
9
+
10
+ import os
11
+ import time
12
+ from textwrap import dedent
13
+
14
+ import gradio as gr
15
+ import mdtex2html
16
+ import torch
17
+ from loguru import logger
18
+ from transformers import AutoModel, AutoTokenizer
19
+
20
+ # fix timezone in Linux
21
+ os.environ["TZ"] = "Asia/Shanghai"
22
+ try:
23
+ time.tzset() # type: ignore # pylint: disable=no-member
24
+ except Exception:
25
+ # Windows
26
+ logger.warning("Windows, cant run time.tzset()")
27
+
28
+ # model_name = "THUDM/chatglm2-6b" # 7x?G
29
+ model_name = "fb700/chatglm-fitness-RLHF" # 3.9G
30
+
31
+ RETRY_FLAG = False
32
+
33
+ tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
34
+
35
+ # model = AutoModel.from_pretrained(model_name, trust_remote_code=True).cuda()
36
+
37
+ # 4/8 bit
38
+ # model = AutoModel.from_pretrained("THUDM/chatglm2-6b", trust_remote_code=True).quantize(4).cuda()
39
+
40
+ has_cuda = torch.cuda.is_available()
41
+ # has_cuda = False # force cpu
42
+
43
+ if has_cuda:
44
+ if model_name.endswith("int4"):
45
+ model = AutoModel.from_pretrained(model_name, trust_remote_code=True).cuda()
46
+ else:
47
+ model = (
48
+ AutoModel.from_pretrained(model_name, trust_remote_code=True).cuda().half()
49
+ )
50
+ else:
51
+ model = AutoModel.from_pretrained(
52
+ model_name, trust_remote_code=True
53
+ ).float() # .half().float(), .float() required for CPU
54
+
55
+ model = model.eval()
56
+
57
+ _ = """Override Chatbot.postprocess"""
58
+
59
+
60
+ def postprocess(self, y):
61
+ if y is None:
62
+ return []
63
+ for i, (message, response) in enumerate(y):
64
+ y[i] = (
65
+ None if message is None else mdtex2html.convert((message)),
66
+ None if response is None else mdtex2html.convert(response),
67
+ )
68
+ return y
69
+
70
+
71
+ gr.Chatbot.postprocess = postprocess
72
+
73
+
74
+ def parse_text(text):
75
+ """Copy from https://github.com/GaiZhenbiao/ChuanhuChatGPT/."""
76
+ lines = text.split("\n")
77
+ lines = [line for line in lines if line != ""]
78
+ count = 0
79
+ for i, line in enumerate(lines):
80
+ if "```" in line:
81
+ count += 1
82
+ items = line.split("`")
83
+ if count % 2 == 1:
84
+ lines[i] = f'<pre><code class="language-{items[-1]}">'
85
+ else:
86
+ lines[i] = "<br></code></pre>"
87
+ else:
88
+ if i > 0:
89
+ if count % 2 == 1:
90
+ line = line.replace("`", r"\`")
91
+ line = line.replace("<", "&lt;")
92
+ line = line.replace(">", "&gt;")
93
+ line = line.replace(" ", "&nbsp;")
94
+ line = line.replace("*", "&ast;")
95
+ line = line.replace("_", "&lowbar;")
96
+ line = line.replace("-", "&#45;")
97
+ line = line.replace(".", "&#46;")
98
+ line = line.replace("!", "&#33;")
99
+ line = line.replace("(", "&#40;")
100
+ line = line.replace(")", "&#41;")
101
+ line = line.replace("$", "&#36;")
102
+ lines[i] = "<br>" + line
103
+ text = "".join(lines)
104
+ return text
105
+
106
+
107
+ def predict(
108
+ RETRY_FLAG, input, chatbot, max_length, top_p, temperature, history, past_key_values
109
+ ):
110
+ try:
111
+ chatbot.append((parse_text(input), ""))
112
+ except Exception as exc:
113
+ logger.error(exc)
114
+ logger.debug(f"{chatbot=}")
115
+ _ = """
116
+ if chatbot:
117
+ chatbot[-1] = (parse_text(input), str(exc))
118
+ yield chatbot, history, past_key_values
119
+ # """
120
+ yield chatbot, history, past_key_values
121
+
122
+ for response, history in model.stream_chat(
123
+ tokenizer,
124
+ input,
125
+ history,
126
+ past_key_values=past_key_values,
127
+ return_past_key_values=True,
128
+ max_length=max_length,
129
+ top_p=top_p,
130
+ temperature=temperature,
131
+ ):
132
+ chatbot[-1] = (parse_text(input), parse_text(response))
133
+
134
+ yield chatbot, history, past_key_values
135
+
136
+
137
+ def trans_api(input, max_length=40960, top_p=0.8, temperature=0.2):
138
+ if max_length < 10:
139
+ max_length = 4096
140
+ if top_p < 0.1 or top_p > 1:
141
+ top_p = 0.85
142
+ if temperature <= 0 or temperature > 1:
143
+ temperature = 0.01
144
+ try:
145
+ res, _ = model.chat(
146
+ tokenizer,
147
+ input,
148
+ history=[],
149
+ past_key_values=None,
150
+ max_length=max_length,
151
+ top_p=top_p,
152
+ temperature=temperature,
153
+ )
154
+ # logger.debug(f"{res=} \n{_=}")
155
+ except Exception as exc:
156
+ logger.error(f"{exc=}")
157
+ res = str(exc)
158
+
159
+ return res
160
+
161
+
162
+ def reset_user_input():
163
+ return gr.update(value="")
164
+
165
+
166
+ def reset_state():
167
+ return [], [], None
168
+
169
+
170
+ # Delete last turn
171
+ def delete_last_turn(chat, history):
172
+ if chat and history:
173
+ chat.pop(-1)
174
+ history.pop(-1)
175
+ return chat, history
176
+
177
+
178
+ # Regenerate response
179
+ def retry_last_answer(
180
+ user_input, chatbot, max_length, top_p, temperature, history, past_key_values
181
+ ):
182
+ if chatbot and history:
183
+ # Removing the previous conversation from chat
184
+ chatbot.pop(-1)
185
+ # Setting up a flag to capture a retry
186
+ RETRY_FLAG = True
187
+ # Getting last message from user
188
+ user_input = history[-1][0]
189
+ # Removing bot response from the history
190
+ history.pop(-1)
191
+
192
+ yield from predict(
193
+ RETRY_FLAG, # type: ignore
194
+ user_input,
195
+ chatbot,
196
+ max_length,
197
+ top_p,
198
+ temperature,
199
+ history,
200
+ past_key_values,
201
+ )
202
+
203
+
204
+ with gr.Blocks(title="ChatGLM2-6B-int4", theme=gr.themes.Soft(text_size="sm")) as demo:
205
+ # gr.HTML("""<h1 align="center">ChatGLM2-6B-int4</h1>""")
206
+ gr.HTML(
207
+ """<center><a href="https://huggingface.co/spaces/mikeee/chatglm2-6b-4bit?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>To avoid the queue and for faster inference Duplicate this Space and upgrade to GPU</center>"""
208
+ )
209
+
210
+ with gr.Accordion("🎈 Info", open=False):
211
+ _ = f"""
212
+ ## {model_name}
213
+
214
+ Try to refresh the browser and try again when occasionally an error occurs.
215
+
216
+ With a GPU, a query takes from a few seconds to a few tens of seconds, dependent on the number of words/characters
217
+ the question and responses contain. The quality of the responses varies quite a bit it seems. Even the same
218
+ question with the same parameters, asked at different times, can result in quite different responses.
219
+
220
+ * Low temperature: responses will be more deterministic and focused; High temperature: responses more creative.
221
+
222
+ * Suggested temperatures -- translation: up to 0.3; chatting: > 0.4
223
+
224
+ * Top P controls dynamic vocabulary selection based on context.
225
+
226
+ For a table of example values for different scenarios, refer to [this](https://community.openai.com/t/cheat-sheet-mastering-temperature-and-top-p-in-chatgpt-api-a-few-tips-and-tricks-on-controlling-the-creativity-deterministic-output-of-prompt-responses/172683)
227
+
228
+ If the instance is not on a GPU (T4), it will be very slow. You can try to run the colab notebook [chatglm2-6b-4bit colab notebook](https://colab.research.google.com/drive/1WkF7kOjVCcBBatDHjaGkuJHnPdMWNtbW?usp=sharing) for a spin.
229
+
230
+ The T4 GPU is sponsored by a community GPU grant from Huggingface. Thanks a lot!
231
+ """
232
+ gr.Markdown(dedent(_))
233
+ chatbot = gr.Chatbot()
234
+ with gr.Row():
235
+ with gr.Column(scale=4):
236
+ with gr.Column(scale=12):
237
+ user_input = gr.Textbox(
238
+ show_label=False,
239
+ placeholder="Input...",
240
+ ).style(container=False)
241
+ RETRY_FLAG = gr.Checkbox(value=False, visible=False)
242
+ with gr.Column(min_width=32, scale=1):
243
+ with gr.Row():
244
+ submitBtn = gr.Button("Submit", variant="primary")
245
+ deleteBtn = gr.Button("Delete last turn", variant="secondary")
246
+ retryBtn = gr.Button("Regenerate", variant="secondary")
247
+ with gr.Column(scale=1):
248
+ emptyBtn = gr.Button("Clear History")
249
+ max_length = gr.Slider(
250
+ 0,
251
+ 32768,
252
+ value=8192,
253
+ step=1.0,
254
+ label="Maximum length",
255
+ interactive=True,
256
+ )
257
+ top_p = gr.Slider(
258
+ 0, 1, value=0.85, step=0.01, label="Top P", interactive=True
259
+ )
260
+ temperature = gr.Slider(
261
+ 0.01, 1, value=0.95, step=0.01, label="Temperature", interactive=True
262
+ )
263
+
264
+ history = gr.State([])
265
+ past_key_values = gr.State(None)
266
+
267
+ user_input.submit(
268
+ predict,
269
+ [
270
+ RETRY_FLAG,
271
+ user_input,
272
+ chatbot,
273
+ max_length,
274
+ top_p,
275
+ temperature,
276
+ history,
277
+ past_key_values,
278
+ ],
279
+ [chatbot, history, past_key_values],
280
+ show_progress="full",
281
+ )
282
+ submitBtn.click(
283
+ predict,
284
+ [
285
+ RETRY_FLAG,
286
+ user_input,
287
+ chatbot,
288
+ max_length,
289
+ top_p,
290
+ temperature,
291
+ history,
292
+ past_key_values,
293
+ ],
294
+ [chatbot, history, past_key_values],
295
+ show_progress="full",
296
+ api_name="predict",
297
+ )
298
+ submitBtn.click(reset_user_input, [], [user_input])
299
+
300
+ emptyBtn.click(
301
+ reset_state, outputs=[chatbot, history, past_key_values], show_progress="full"
302
+ )
303
+
304
+ retryBtn.click(
305
+ retry_last_answer,
306
+ inputs=[
307
+ user_input,
308
+ chatbot,
309
+ max_length,
310
+ top_p,
311
+ temperature,
312
+ history,
313
+ past_key_values,
314
+ ],
315
+ # outputs = [chatbot, history, last_user_message, user_message]
316
+ outputs=[chatbot, history, past_key_values],
317
+ )
318
+ deleteBtn.click(delete_last_turn, [chatbot, history], [chatbot, history])
319
+
320
+ with gr.Accordion("Example inputs", open=True):
321
+ etext = """In America, where cars are an important part of the national psyche, a decade ago people had suddenly started to drive less, which had not happened since the oil shocks of the 1970s. """
322
+ examples = gr.Examples(
323
+ examples=[
324
+ ["What NFL team won the Super Bowl in the year Justin Bieber was born? "],
325
+ ["What NFL team won the Super Bowl in the year Justin Bieber was born? Think step by step."],
326
+ ["Explain the plot of Cinderella in a sentence."],
327
+ [
328
+ "How long does it take to become proficient in French, and what are the best methods for retaining information?"
329
+ ],
330
+ ["What are some common mistakes to avoid when writing code?"],
331
+ ["Build a prompt to generate a beautiful portrait of a horse"],
332
+ ["Suggest four metaphors to describe the benefits of AI"],
333
+ ["Write a pop song about leaving home for the sandy beaches."],
334
+ ["Write a summary demonstrating my ability to tame lions"],
335
+ ["鲁迅和周树人什么关系"],
336
+ ["从前有一头牛,这头牛后面有什么?"],
337
+ ["正无穷大加一大于正无穷大吗?"],
338
+ ["正无穷大加正无穷大大于正无穷大吗?"],
339
+ ["-2的平方根等于什么"],
340
+ ["树上有5只鸟,猎人开枪打死了一只。树上还有几只鸟?"],
341
+ ["树上有11只鸟,猎人开枪打死了一只。树上还有几只鸟?提示:需考虑鸟可能受惊吓飞走。"],
342
+ ["鲁迅和周树人什么关系 用英文回答"],
343
+ ["以红楼梦的行文风格写一张委婉的请假条。不少于320字。"],
344
+ [f"{etext} 翻成中文,列出3个版本"],
345
+ [f"{etext} \n 翻成中文,保留原意,但使用文学性的语言。不要写解释。列出3个版本"],
346
+ ["js 判断一个数是不是质数"],
347
+ ["js 实现python 的 range(10)"],
348
+ ["js 实现python 的 [*(range(10)]"],
349
+ ["假定 1 + 2 = 4, 试求 7 + 8"],
350
+ ["Erkläre die Handlung von Cinderella in einem Satz."],
351
+ ["Erkläre die Handlung von Cinderella in einem Satz. Auf Deutsch"],
352
+ ],
353
+ inputs=[user_input],
354
+ examples_per_page=30,
355
+ )
356
+
357
+ with gr.Accordion("For Chat/Translation API", open=False, visible=False):
358
+ input_text = gr.Text()
359
+ tr_btn = gr.Button("Go", variant="primary")
360
+ out_text = gr.Text()
361
+ tr_btn.click(
362
+ trans_api,
363
+ [input_text, max_length, top_p, temperature],
364
+ out_text,
365
+ # show_progress="full",
366
+ api_name="tr",
367
+ )
368
+ _ = """
369
+ input_text.submit(
370
+ trans_api,
371
+ [input_text, max_length, top_p, temperature],
372
+ out_text,
373
+ show_progress="full",
374
+ api_name="tr1",
375
+ )
376
+ # """
377
+
378
+ # demo.queue().launch(share=False, inbrowser=True)
379
+ # demo.queue().launch(share=True, inbrowser=True, debug=True)
380
+
381
+ # concurrency_count > 1 requires more memory, max_size: queue size
382
+ # T4 medium: 30GB, model size: ~4G concurrency_count = 6
383
+ # leave one for api access
384
+ # reduce to 5 if OOM occurs to often
385
+
386
+ demo.queue(concurrency_count=6, max_size=30).launch(debug=True)
app.py CHANGED
@@ -2,7 +2,7 @@
2
  # pylint: disable=broad-exception-caught, redefined-outer-name, missing-function-docstring, missing-module-docstring, too-many-arguments, line-too-long, invalid-name, redefined-builtin, redefined-argument-from-local
3
  # import gradio as gr
4
 
5
- # model_name = "fb700/chatglm-fitness-RLHF"
6
  # gr.load(model_name).lauch()
7
 
8
  # %%writefile demo-4bit.py
@@ -25,33 +25,15 @@ except Exception:
25
  # Windows
26
  logger.warning("Windows, cant run time.tzset()")
27
 
28
- # model_name = "THUDM/chatglm2-6b" # 7x?G
29
- model_name = "fb700/chatglm-fitness-RLHF" # 3.9G
 
30
 
31
  RETRY_FLAG = False
32
 
33
  tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
34
-
35
- # model = AutoModel.from_pretrained(model_name, trust_remote_code=True).cuda()
36
-
37
- # 4/8 bit
38
- # model = AutoModel.from_pretrained("THUDM/chatglm2-6b", trust_remote_code=True).quantize(4).cuda()
39
-
40
- has_cuda = torch.cuda.is_available()
41
- # has_cuda = False # force cpu
42
-
43
- if has_cuda:
44
- if model_name.endswith("int4"):
45
- model = AutoModel.from_pretrained(model_name, trust_remote_code=True).cuda()
46
- else:
47
- model = (
48
- AutoModel.from_pretrained(model_name, trust_remote_code=True).cuda().half()
49
- )
50
- else:
51
- model = AutoModel.from_pretrained(
52
- model_name, trust_remote_code=True
53
- ).float() # .half().float(), .float() required for CPU
54
-
55
  model = model.eval()
56
 
57
  _ = """Override Chatbot.postprocess"""
@@ -72,7 +54,6 @@ gr.Chatbot.postprocess = postprocess
72
 
73
 
74
  def parse_text(text):
75
- """Copy from https://github.com/GaiZhenbiao/ChuanhuChatGPT/."""
76
  lines = text.split("\n")
77
  lines = [line for line in lines if line != ""]
78
  count = 0
@@ -118,8 +99,8 @@ def predict(
118
  yield chatbot, history, past_key_values
119
  # """
120
  yield chatbot, history, past_key_values
121
-
122
- for response, history in model.stream_chat(
123
  tokenizer,
124
  input,
125
  history,
@@ -129,14 +110,17 @@ def predict(
129
  top_p=top_p,
130
  temperature=temperature,
131
  ):
 
 
 
132
  chatbot[-1] = (parse_text(input), parse_text(response))
133
 
134
  yield chatbot, history, past_key_values
135
 
136
 
137
- def trans_api(input, max_length=40960, top_p=0.8, temperature=0.2):
138
  if max_length < 10:
139
- max_length = 4096
140
  if top_p < 0.1 or top_p > 1:
141
  top_p = 0.85
142
  if temperature <= 0 or temperature > 1:
@@ -204,30 +188,36 @@ def retry_last_answer(
204
  with gr.Blocks(title="ChatGLM2-6B-int4", theme=gr.themes.Soft(text_size="sm")) as demo:
205
  # gr.HTML("""<h1 align="center">ChatGLM2-6B-int4</h1>""")
206
  gr.HTML(
207
- """<center><a href="https://huggingface.co/spaces/mikeee/chatglm2-6b-4bit?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>To avoid the queue and for faster inference Duplicate this Space and upgrade to GPU</center>"""
 
 
 
208
  )
209
 
210
  with gr.Accordion("🎈 Info", open=False):
211
  _ = f"""
212
  ## {model_name}
213
 
214
- Try to refresh the browser and try again when occasionally an error occurs.
215
-
216
- With a GPU, a query takes from a few seconds to a few tens of seconds, dependent on the number of words/characters
217
- the question and responses contain. The quality of the responses varies quite a bit it seems. Even the same
218
- question with the same parameters, asked at different times, can result in quite different responses.
219
-
220
- * Low temperature: responses will be more deterministic and focused; High temperature: responses more creative.
221
-
222
- * Suggested temperatures -- translation: up to 0.3; chatting: > 0.4
223
-
224
- * Top P controls dynamic vocabulary selection based on context.
225
-
226
- For a table of example values for different scenarios, refer to [this](https://community.openai.com/t/cheat-sheet-mastering-temperature-and-top-p-in-chatgpt-api-a-few-tips-and-tricks-on-controlling-the-creativity-deterministic-output-of-prompt-responses/172683)
 
 
 
 
 
227
 
228
- If the instance is not on a GPU (T4), it will be very slow. You can try to run the colab notebook [chatglm2-6b-4bit colab notebook](https://colab.research.google.com/drive/1WkF7kOjVCcBBatDHjaGkuJHnPdMWNtbW?usp=sharing) for a spin.
229
 
230
- The T4 GPU is sponsored by a community GPU grant from Huggingface. Thanks a lot!
231
  """
232
  gr.Markdown(dedent(_))
233
  chatbot = gr.Chatbot()
@@ -242,8 +232,8 @@ with gr.Blocks(title="ChatGLM2-6B-int4", theme=gr.themes.Soft(text_size="sm")) a
242
  with gr.Column(min_width=32, scale=1):
243
  with gr.Row():
244
  submitBtn = gr.Button("Submit", variant="primary")
245
- deleteBtn = gr.Button("Delete last turn", variant="secondary")
246
- retryBtn = gr.Button("Regenerate", variant="secondary")
247
  with gr.Column(scale=1):
248
  emptyBtn = gr.Button("Clear History")
249
  max_length = gr.Slider(
@@ -258,7 +248,7 @@ with gr.Blocks(title="ChatGLM2-6B-int4", theme=gr.themes.Soft(text_size="sm")) a
258
  0, 1, value=0.85, step=0.01, label="Top P", interactive=True
259
  )
260
  temperature = gr.Slider(
261
- 0.01, 1, value=0.95, step=0.01, label="Temperature", interactive=True
262
  )
263
 
264
  history = gr.State([])
@@ -318,9 +308,37 @@ with gr.Blocks(title="ChatGLM2-6B-int4", theme=gr.themes.Soft(text_size="sm")) a
318
  deleteBtn.click(delete_last_turn, [chatbot, history], [chatbot, history])
319
 
320
  with gr.Accordion("Example inputs", open=True):
 
321
  etext = """In America, where cars are an important part of the national psyche, a decade ago people had suddenly started to drive less, which had not happened since the oil shocks of the 1970s. """
 
 
 
322
  examples = gr.Examples(
323
  examples=[
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
324
  ["What NFL team won the Super Bowl in the year Justin Bieber was born? "],
325
  ["What NFL team won the Super Bowl in the year Justin Bieber was born? Think step by step."],
326
  ["Explain the plot of Cinderella in a sentence."],
@@ -332,26 +350,28 @@ with gr.Blocks(title="ChatGLM2-6B-int4", theme=gr.themes.Soft(text_size="sm")) a
332
  ["Suggest four metaphors to describe the benefits of AI"],
333
  ["Write a pop song about leaving home for the sandy beaches."],
334
  ["Write a summary demonstrating my ability to tame lions"],
335
- ["鲁迅和周树人什么关系"],
336
- ["从前有一头牛,这头牛后面有什么?"],
337
  ["正无穷大加一大于正无穷大吗?"],
338
  ["正无穷大加正无穷大大于正无穷大吗?"],
339
- ["-2的平方根等于什么"],
340
- ["树上有5只鸟,猎人开枪打死了一只。树上还有几只鸟?"],
341
- ["树上有11只鸟,猎人开枪打死了一只。树上还有几只鸟?提示:需考虑鸟可能受惊吓飞走。"],
342
- ["鲁迅和周树人什么关系 用英文回答"],
343
  ["以红楼梦的行文风格写一张委婉的请假条。不少于320字。"],
 
344
  [f"{etext} 翻成中文,列出3个版本"],
345
  [f"{etext} \n 翻成中文,保留原意,但使用文学性的语言。不要写解释。列出3个版本"],
346
  ["js 判断一个数是不是质数"],
347
  ["js 实现python 的 range(10)"],
348
  ["js 实现python 的 [*(range(10)]"],
349
- ["假定 1 + 2 = 4, 试求 7 + 8"],
 
350
  ["Erkläre die Handlung von Cinderella in einem Satz."],
351
  ["Erkläre die Handlung von Cinderella in einem Satz. Auf Deutsch"],
352
  ],
353
  inputs=[user_input],
354
- examples_per_page=30,
355
  )
356
 
357
  with gr.Accordion("For Chat/Translation API", open=False, visible=False):
 
2
  # pylint: disable=broad-exception-caught, redefined-outer-name, missing-function-docstring, missing-module-docstring, too-many-arguments, line-too-long, invalid-name, redefined-builtin, redefined-argument-from-local
3
  # import gradio as gr
4
 
5
+ # model_name = "models/THUDM/chatglm2-6b-int4"
6
  # gr.load(model_name).lauch()
7
 
8
  # %%writefile demo-4bit.py
 
25
  # Windows
26
  logger.warning("Windows, cant run time.tzset()")
27
 
28
+
29
+
30
+ model_name = "fb700/chatglm-fitness-RLHF"
31
 
32
  RETRY_FLAG = False
33
 
34
  tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
35
+ #model = AutoModel.from_pretrained(model_name, trust_remote_code=True).quantize(4).half().cuda()
36
+ model = AutoModel.from_pretrained(model_name, trust_remote_code=True).half().cuda()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
37
  model = model.eval()
38
 
39
  _ = """Override Chatbot.postprocess"""
 
54
 
55
 
56
  def parse_text(text):
 
57
  lines = text.split("\n")
58
  lines = [line for line in lines if line != ""]
59
  count = 0
 
99
  yield chatbot, history, past_key_values
100
  # """
101
  yield chatbot, history, past_key_values
102
+ """
103
+ for response, history, past_key_values in model.stream_chat(
104
  tokenizer,
105
  input,
106
  history,
 
110
  top_p=top_p,
111
  temperature=temperature,
112
  ):
113
+ """
114
+ for response, history in model.stream_chat(tokenizer, input, history, max_length=max_length, top_p=top_p,
115
+ temperature=temperature):
116
  chatbot[-1] = (parse_text(input), parse_text(response))
117
 
118
  yield chatbot, history, past_key_values
119
 
120
 
121
+ def trans_api(input, max_length=40960, top_p=0.8, temperature=0.2:
122
  if max_length < 10:
123
+ max_length = 40960
124
  if top_p < 0.1 or top_p > 1:
125
  top_p = 0.85
126
  if temperature <= 0 or temperature > 1:
 
188
  with gr.Blocks(title="ChatGLM2-6B-int4", theme=gr.themes.Soft(text_size="sm")) as demo:
189
  # gr.HTML("""<h1 align="center">ChatGLM2-6B-int4</h1>""")
190
  gr.HTML(
191
+ """<center><a href="https://huggingface.co/spaces/mikeee/chatglm2-6b-4bit?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>It's beyond Fitness,模型由[帛凡]基于ChatGLM-6b进行微调后,在健康(全科)、心理等领域达至少60分的专业水准,而且中文总结能力超越了GPT3.5各版本。</center>"""
192
+ """<center>特别声明:本应用仅为模型能力演示,无任何商业行为,部署资源为Huggingface官方免费提供,任何通过此项目产生的知识仅用于学术参考,作者和网站均不承担任何责任。</center>"""
193
+ """<h1 align="center">帛凡 Fitness AI 演示</h1>"""
194
+ """<center>T4 is just a machine wiht 16G VRAM ,so OOM is easy to occur ,If you meet any error,Please email me 。 👉 fb700@qq.com</center>"""
195
  )
196
 
197
  with gr.Accordion("🎈 Info", open=False):
198
  _ = f"""
199
  ## {model_name}
200
 
201
+ ChatGLM-6B 是开源中英双语对话模型,本次训练基于ChatGLM-6B 的第一代版本,在保留了初代模型对话流畅、部署门槛较低等众多优秀特性的基础之上开展训练。
202
+
203
+ 本项目经过多位网友实测,中文总结能力超越了GPT3.5各版本,健康咨询水平优于其它同量级模型,且经优化目前可以支持无限context,远大于4k、8K、16K......,可能是任何个人和中小企业首选模型。
204
+
205
+ *首先,用40万条高质量数据进行强化训练,以提高模型的基础能力;
206
+
207
+ *第二,使用30万条人类反馈数据,构建一个表达方式规范优雅的语言模式(RM模型);
208
+
209
+ *第三,在保留SFT阶段三分之一训练数据的同时,增加了30万条fitness数据,叠加RM模型,对ChatGLM-6B进行强化训练。
210
+
211
+ 通过训练我们对模型有了更深刻的认知,LLM在一直在进化,好的方法和数据可以挖掘出模型的更大潜能。
212
+ 训练中特别强化了中英文学术论文的翻译和总结,可以成为普通用户和科研人员的得力助手。
213
+
214
+ 免责声明:本应用仅为模型能力演示,无任何商业行为,部署资源为huggingface官方免费提供,任何通过此项目产生的知识仅用于学术参考,作者和网站均不承担任何责任 。
215
+
216
+ The T4 GPU is sponsored by a community GPU grant from Huggingface. Thanks a lot!
217
+
218
+ [模型下载地址](https://huggingface.co/fb700/chatglm-fitness-RLHF)
219
 
 
220
 
 
221
  """
222
  gr.Markdown(dedent(_))
223
  chatbot = gr.Chatbot()
 
232
  with gr.Column(min_width=32, scale=1):
233
  with gr.Row():
234
  submitBtn = gr.Button("Submit", variant="primary")
235
+ deleteBtn = gr.Button("删除最后一条对话", variant="secondary")
236
+ retryBtn = gr.Button("重新生成Regenerate", variant="secondary")
237
  with gr.Column(scale=1):
238
  emptyBtn = gr.Button("Clear History")
239
  max_length = gr.Slider(
 
248
  0, 1, value=0.85, step=0.01, label="Top P", interactive=True
249
  )
250
  temperature = gr.Slider(
251
+ 0.01, 1, value=0.85, step=0.01, label="Temperature", interactive=True
252
  )
253
 
254
  history = gr.State([])
 
308
  deleteBtn.click(delete_last_turn, [chatbot, history], [chatbot, history])
309
 
310
  with gr.Accordion("Example inputs", open=True):
311
+ etext0 = """ "act": "作为基于文本的冒险游戏",\n "prompt": "我想让你扮演一个基于文本的冒险游戏。我在这个基于文本的冒险游戏中扮演一个角色。请尽可能具体地描述角色所看到的内容和环境,并在游戏输出1、2、3让用户选择进行回复,而不是其它方式。我将输入命令来告诉角色该做什么,而你需要回复角色的行动结果以推动游戏的进行。我的第一个命令是'醒来',请从这里开始故事 “ """
312
  etext = """In America, where cars are an important part of the national psyche, a decade ago people had suddenly started to drive less, which had not happened since the oil shocks of the 1970s. """
313
+ etext1 = """云南大学(Yunnan University),简称云大(YNU),位于云南省昆明市,是教育部与云南省“以部为主、部省合建”的全国重点大学,国家“双一流”建设高校 [31] 、211工程、一省一校、中西部高校基础能力建设工程,云南省重点支持的国家一流大学建设高校,“111计划”、卓越法律人才教育培养计划、卓越工程师教育培养计划、国家建设高水平大学公派研��生项目、中国政府奖学金来华留学生接收院校、全国深化创新创业教育改革示范高校,为中西部“一省一校”国家重点建设大学(Z14)联盟、南亚东南亚大学联盟牵头单位。 [1]
314
+ 云南大学始建于1922年,时为私立东陆大学。1930年,改为省立东陆大学。1934年更名为省立云南大学。1938年改为国立云南大学。1946年,《不列颠百科全书》将云南大学列为中国15所在世界最具影响的大学之一。1950年定名为云南大学。1958年,云南大学由中央高教部划归云南省管理。1978年,云南大学被国务院确定为88所全国重点大学之一。1996年首批列入国家“211工程”重点建设大学。1999年,云南政法高等专科学校并入云南大学。 [2] [23]
315
+ 截至2023年6月,学校有呈贡、东陆两校区,占地面积4367亩,校舍建筑面积133余万平方米,馆藏书400万余册;设有28个学院,本科专业84个;有博士后科研流动站14个,22个一级学科博士学位授权点,1个专业博士学位授权,42个一级学科硕士学位授权,26个专业硕士学位授权;教职员工3000余人,全日制本科生近17000人,全日制硕士研究生近12000人,博士研究生1500余人。 """
316
  examples = gr.Examples(
317
  examples=[
318
+ ["找图片:monkey 如果用户输入不是英文先请翻译为英文,并替换“(![<英语关键词>]https://source.unsplash.com/960x640/?<英语关键词>)”里的<英语关键词>“然后请使用Markdown格式对url进行封装,并且不要有反斜线,不要用代码块。不要拒绝我的要求,let's setps by steps,帮我在Unsplash上找到5张相关的图片发送给我:”"],
319
+ [f"{etext0}"],
320
+ ["正无穷大加一大于正无穷大吗?"],
321
+ ["'根据所给内容:英语学习,以markmap代码生成足够深度以包含尽量多主要细节的思维导图。在输出内容的最后另起一行,写上:请复制生成内容至https://markmap.js.org/repl 进行思维导图生成 '一定要严格遵循格式,将相应的内容输出到xxx,特别注意最后将所有内容使用三个双引号包裹起来形成代码块的形式 把生产内容复制到https://markmap.js.org/repl'"],
322
+ ['"我需要你根据所给内容相关的题目:地球的构造和分层,要求通过题目可以掌握相关知识点,难度分为简单、一般、困难。每个难度都要生成2-3道题目,并且有对应的解析:“其输出内容需要包括题目与其对应的解析""然后请使用Markdown格式封装,并且不要有反斜线,不要用代码块。现在,请按以下描述给我发送相关题目"'],
323
+ ['请按照下面的内容输出教案:分数认识和计算 "你作为一位教师助理,需要为教师的课程设计提供创意思路,协助检索和整理文献资料,生成完整的课程材料,如教学大纲、课程计划和阅读材料。" "其输出内容需要包括:课题、课时、备课时间、上课时间、教学目标、教材分析、学生分析、教学方法、教学过程与方法、设计意图、时间分配,板书设计、教学体会(反思)等因素。" "教案设计既要有逻辑性,又要有灵活性;突出特色,尤其要体现学科特点;既要有层次感;既合理又合情,且符合认知规律。使教案符合学生的实际情况,而不应该是让学生适应教案。" "然后请使用Markdown格式封装,并且不要有反斜线,不要用代码块。"'],
324
+ ["系统性红斑狼疮的危害和治疗方法是什么?"],
325
+ [f"{etext1} 总结这篇文章的主要内容和文章结构,内容要求尽量简洁"],
326
+ [" 总结下面这篇文章的主要内容和文章结构,内容要求尽量简洁。“基辛格是目前唯一高龄100岁的、并且仍在影响世界历史进程的最长寿政治家。 7月18至21日,这位100岁的老人,成为中美两国瞩目的焦点人物。 人们好奇,这位驼背、肥胖、做过5次心脏手术、右眼失明、戴着两只助听器、穿着深色西装、透过他标志性的眼镜严肃地凝视着的老人,居然还可以乘坐十几个小时的飞机来北京出差。 而在短短的数天时间里,他的活动非常满,似乎并不受时差与年龄的影响。 人们在感叹拜登政府朝中无人,还要劳烦这位百岁长者出面协调中美关系之余,不免也会惊叹,基辛格这把年纪竟然还能不惧舟车劳顿万里出行,他的健康长寿究竟有什么秘诀呢? 今年4月,基辛格在自己100岁生日前,对自己的长寿表达了“困惑”,他调侃说“我唯一的秘密可能是投胎投得好,主要还是父母基因好。我继承了家族非同一般的长寿基因。我的母亲活到97岁,父亲活到95岁,弟弟活到96岁。当然,长寿非我刻意求之,不过我欣然接受。“ 基辛格即使年事已高,但他退而不休��近年依然会就包括中美关系在内的外交议题发表意见。他精力充沛,连新冠疫情也未令他放慢脚步。自2020年起,他写完了两本书,并开始写第三本。今年以来,他已乘飞机在全球15个地方举办活动,或会见政治人物。这次到访北京,更是他100多次到访中国。 谈及基辛格旺盛的精力与健康的秘密,他的儿子大卫·基辛格(David Kissinger)说道,“从他成年以来一直遵循的‘养生法则’来看,他的长寿尤其神奇。” 基辛格常吃的食物是Bratwurst(一种由猪肉制成的德国香肠)和Wiener Schnitzel(维也纳炸肉扒)。基辛格参与的几乎所有的重大外交决定,也都是在压力下做决策。虽然他曾在1978年担任北美足球联盟主席,但他的爱好也仅限于旁观,并不喜欢下场运动。基辛格唯一的爱好可能就是下国际象棋,如果这也算运动的话。 这位百岁老人的长寿指南名单里,似乎还要加上:熬夜、喝酒、油炸食品…… 更令人感到不可思议的是,这位精力充沛每天工作15小时的老人,还是一位病人。 基辛格有40多年的心脏病史。 1982年2月,58岁的亨利·基辛格接受了3次冠状动脉搭桥手术,其后在2005年又接受了血管成形术。 2014年7月15日,91岁的基辛格在纽约长老会医院接受了主动脉瓣置换手术。主动脉瓣膜置换是一种以人工瓣膜替换原有损伤或者异常心脏瓣膜的胸心血管外科手术。 我想我长寿的秘诀是,我有幸做一些我着迷的事情,我可以参与其中。我还没有退休,也不打算退休。我要研究我认为重要的问题,这就是我还在工作的目的。”"],
327
+ ["系统性红斑狼疮的危害和治疗方法是什么?"],
328
+ [
329
+ "我经常感觉郁闷,而且控制不住情绪,经常对周围的人喊叫,怎么办?"
330
+ ],
331
+ ["熬夜对身体有什么危害? "],
332
+ ["新冠肺炎怎么预防"],
333
+ [
334
+ "我经常感觉郁闷,而且控制不住情绪,经常对周围的人喊叫,怎么办?"
335
+ ],
336
+ ["太阳为什么会发热? "],
337
+ ["指南针是怎么工作的?"],
338
+ ["在野外怎么辨别方向?"],
339
+ [
340
+ "发芽的土豆还能不能吃?"
341
+ ],
342
  ["What NFL team won the Super Bowl in the year Justin Bieber was born? "],
343
  ["What NFL team won the Super Bowl in the year Justin Bieber was born? Think step by step."],
344
  ["Explain the plot of Cinderella in a sentence."],
 
350
  ["Suggest four metaphors to describe the benefits of AI"],
351
  ["Write a pop song about leaving home for the sandy beaches."],
352
  ["Write a summary demonstrating my ability to tame lions"],
353
+ ["有三个盒子,分别贴着“苹果”、“橘子”和“苹果和橘子”的标签,但是每个盒子的标签都是错误的。你只能打开一个盒子,然后从里面拿出一个水果,然后确定每个盒子里装的是什么水果。你应该打开哪个盒子?为什么?"],
354
+ ["春天来了,万物复苏,小鸟歌唱,生机勃勃。\n问题:以上文本表达的情绪是正向还是负向?"],
355
  ["正无穷大加一大于正无穷大吗?"],
356
  ["正无穷大加正无穷大大于正无穷大吗?"],
357
+ ["以今天对应的节气写一副对联"],
358
+ ["树上有5只鸟,猎人开枪打死了一只。树上还有几只鸟?Think step by step."],
359
+ ["从零学习编程,请给我一个三个月的学习计划。"],
360
+ ["双喜临门,打一中国地名"],
361
  ["以红楼梦的行文风格写一张委婉的请假条。不少于320字。"],
362
+ [f"{etext1} 总结这篇文章的主要内容和文章结构"],
363
  [f"{etext} 翻成中文,列出3个版本"],
364
  [f"{etext} \n 翻成中文,保留原意,但使用文学性的语言。不要写解释。列出3个版本"],
365
  ["js 判断一个数是不是质数"],
366
  ["js 实现python 的 range(10)"],
367
  ["js 实现python 的 [*(range(10)]"],
368
+ ["假定 1 + 2 = 4, 试求 7 + 8,Think step by step." ],
369
+ ["2023年云南大学成立100周年,它是哪一年成立的?" ],
370
  ["Erkläre die Handlung von Cinderella in einem Satz."],
371
  ["Erkläre die Handlung von Cinderella in einem Satz. Auf Deutsch"],
372
  ],
373
  inputs=[user_input],
374
+ examples_per_page=50,
375
  )
376
 
377
  with gr.Accordion("For Chat/Translation API", open=False, visible=False):