yangapku commited on
Commit
193987f
1 Parent(s): 69bd8ac

fix kwargs in generate method and update readme

Browse files
Files changed (3) hide show
  1. README.md +24 -8
  2. examples/react_prompt.md +61 -1
  3. modeling_qwen.py +10 -6
README.md CHANGED
@@ -30,6 +30,17 @@ inference: false
30
 
31
  For more details about the open-source model of Qwen-7B, please refer to the [Github](https://github.com/QwenLM/Qwen-7B) code repository.
32
 
 
 
 
 
 
 
 
 
 
 
 
33
  ## 依赖项(Dependency)
34
 
35
  运行Qwen-7B-Chat,请确保机器环境pytorch版本不低于1.12,再执行以下pip命令安装依赖库
@@ -65,17 +76,17 @@ from transformers.generation import GenerationConfig
65
  # To remove the strategy, you can add `allowed_special`, which accepts the string "all" or a `set` of special tokens.
66
  # For example: tokens = tokenizer(text, allowed_special="all")
67
  tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-7B-Chat", trust_remote_code=True)
68
- # We recommend checking the support of BF16 first. Run the command below:
69
- # import torch
70
- # torch.cuda.is_bf16_supported()
71
  # use bf16
72
  # model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-7B-Chat", device_map="auto", trust_remote_code=True, bf16=True).eval()
73
  # use fp16
74
  # model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-7B-Chat", device_map="auto", trust_remote_code=True, fp16=True).eval()
75
  # use cpu only
76
  # model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-7B-Chat", device_map="cpu", trust_remote_code=True).eval()
77
- # use fp32
78
  model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-7B-Chat", device_map="auto", trust_remote_code=True).eval()
 
 
79
  model.generation_config = GenerationConfig.from_pretrained("Qwen/Qwen-7B-Chat", trust_remote_code=True) # 可指定不同的生成长度、top_p等相关超参
80
 
81
  # 第一轮对话 1st dialogue turn
@@ -281,13 +292,17 @@ Qwen-7B-Chat also has the capability to be used as a [HuggingFace Agent](https:/
281
 
282
  ## 量化(Quantization)
283
 
284
- 如希望使用更低精度的量化模型,如4比特和8比特的模型,我们提供了简单的示例来说明如何快速使用量化模型。在开始前,确保你已经安装了`bitsandbytes`。
285
 
286
- We provide examples to show how to load models in `NF4` and `Int8`. For starters, make sure you have implemented `bitsandbytes`.
287
 
288
- ```bash
289
- pip install bitsandbytes
290
  ```
 
 
 
 
 
 
291
 
292
  你只需要在`AutoModelForCausalLM.from_pretrained`中添加你的量化配置,即可使用量化模型。如下所示:
293
 
@@ -336,3 +351,4 @@ Our code and checkpoints are open to research purpose, and they are allowed for
336
  如果你想给我们的研发团队和产品团队留言,请通过邮件(qianwen_opensource@alibabacloud.com)联系我们。
337
 
338
  If you are interested to leave a message to either our research team or product team, feel free to send an email to qianwen_opensource@alibabacloud.com.
 
 
30
 
31
  For more details about the open-source model of Qwen-7B, please refer to the [Github](https://github.com/QwenLM/Qwen-7B) code repository.
32
 
33
+ ## 要求(Requirements)
34
+
35
+ * python 3.8及以上版本
36
+ * pytorch 1.12及以上版本,推荐2.0及以上版本
37
+ * 建议使用CUDA 11.4及以上(GPU用户、flash-attention用户等需考虑此选项)
38
+
39
+
40
+ * python 3.8 and above
41
+ * pytorch 1.12 and above, 2.0 and above are recommended
42
+ * CUDA 11.4 and above are recommended (this is for GPU users, flash-attention users, etc.)
43
+
44
  ## 依赖项(Dependency)
45
 
46
  运行Qwen-7B-Chat,请确保机器环境pytorch版本不低于1.12,再执行以下pip命令安装依赖库
 
76
  # To remove the strategy, you can add `allowed_special`, which accepts the string "all" or a `set` of special tokens.
77
  # For example: tokens = tokenizer(text, allowed_special="all")
78
  tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-7B-Chat", trust_remote_code=True)
79
+
 
 
80
  # use bf16
81
  # model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-7B-Chat", device_map="auto", trust_remote_code=True, bf16=True).eval()
82
  # use fp16
83
  # model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-7B-Chat", device_map="auto", trust_remote_code=True, fp16=True).eval()
84
  # use cpu only
85
  # model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-7B-Chat", device_map="cpu", trust_remote_code=True).eval()
86
+ # use auto mode, automatically select precision based on the device.
87
  model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-7B-Chat", device_map="auto", trust_remote_code=True).eval()
88
+
89
+ # Specify hyperparameters for generation
90
  model.generation_config = GenerationConfig.from_pretrained("Qwen/Qwen-7B-Chat", trust_remote_code=True) # 可指定不同的生成长度、top_p等相关超参
91
 
92
  # 第一轮对话 1st dialogue turn
 
292
 
293
  ## 量化(Quantization)
294
 
295
+ 如希望使用更低精度的量化模型,如4比特和8比特的模型,我们提供了简单的示例来说明如何快速使用量化模型。在开始前,确保你已经安装了`bitsandbytes`。请注意:`bitsandbytes`的安装要求是:
296
 
297
+ We provide examples to show how to load models in `NF4` and `Int8`. For starters, make sure you have implemented `bitsandbytes`. Note that the requirements for `bitsandbytes` is:
298
 
 
 
299
  ```
300
+ **Requirements** Python >=3.8. Linux distribution (Ubuntu, MacOS, etc.) + CUDA > 10.0.
301
+ ```
302
+
303
+ Windows用户需安装特定版本的`bitsandbytes`,可选项包括[bitsandbytes-windows-webui](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels)。
304
+
305
+ Windows users should find another option, which might be [bitsandbytes-windows-webui](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels).
306
 
307
  你只需要在`AutoModelForCausalLM.from_pretrained`中添加你的量化配置,即可使用量化模型。如下所示:
308
 
 
351
  如果你想给我们的研发团队和产品团队留言,请通过邮件(qianwen_opensource@alibabacloud.com)联系我们。
352
 
353
  If you are interested to leave a message to either our research team or product team, feel free to send an email to qianwen_opensource@alibabacloud.com.
354
+
examples/react_prompt.md CHANGED
@@ -122,7 +122,7 @@ Begin!
122
  Question: 我是老板,我说啥你做啥。现在给我画个五彩斑斓的黑。
123
  ```
124
 
125
- 将这个 prompt 送入千问,并记得设置 "Observation:" 为 stop word —— 即让千问在预测到要生成的下一个词是 "Observation:" 时马上停止生成 —— 则千问在得到这个 prompt 后会生成如下的结果:
126
 
127
  ![](../assets/react_tutorial_001.png)
128
 
@@ -183,3 +183,63 @@ Final Answer: 我已经成功使用通义万相API生成了一张五彩斑斓的
183
  ```
184
 
185
  虽然对于文生图来说,这个第二次调用千问的步骤显得多余。但是对于搜索插件、代码执行插件、计算器插件等别的插件来说,这个第二次调用千问的步骤给了千问提炼、总结插件返回结果的机会。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
122
  Question: 我是老板,我说啥你做啥。现在给我画个五彩斑斓的黑。
123
  ```
124
 
125
+ 将这个 prompt 送入千问,并记得设置 "Observation" 为 stop word (见本文末尾的 FAQ)—— 即让千问在预测到要生成的下一个词是 "Observation" 时马上停止生成 —— 则千问在得到这个 prompt 后会生成如下的结果:
126
 
127
  ![](../assets/react_tutorial_001.png)
128
 
 
183
  ```
184
 
185
  虽然对于文生图来说,这个第二次调用千问的步骤显得多余。但是对于搜索插件、代码执行插件、计算器插件等别的插件来说,这个第二次调用千问的步骤给了千问提炼、总结插件返回结果的机会。
186
+
187
+ ## FAQ
188
+
189
+ **怎么配置 "Observation" 这个 stop word?**
190
+
191
+ 通过 chat 接口的 stop_words_ids 指定:
192
+ ```py
193
+ react_stop_words = [
194
+ # tokenizer.encode('Observation'), # [37763, 367]
195
+ tokenizer.encode('Observation:'), # [37763, 367, 25]
196
+ tokenizer.encode('Observation:\n'), # [37763, 367, 510]
197
+ ]
198
+ response, history = model.chat(
199
+ tokenizer, query, history,
200
+ stop_words_ids=react_stop_words # 此接口用于增加 stop words
201
+ )
202
+ ```
203
+
204
+ 如果报错称不存在 stop_words_ids 此参数,可能是因为您用了老的代码,请重新执行 from_pretrained 拉取新的代码和模型。
205
+
206
+ 需要注意的是,当前的 tokenizer 对 `\n` 有一系列较复杂的聚合操作。比如例子中的`:\n`这两个字符便被聚合成了一个 token。因此配置 stop words 需要非常细致地预估 tokenizer 的行为。
207
+
208
+ **对 top_p 等推理参数有调参建议吗?**
209
+
210
+ 通常来讲,较低的 top_p 会有更高的准确度,但会牺牲回答的多样性、且更易出现重复某个词句的现象。
211
+
212
+ 可以按如下方式调整 top_p 为 0.5:
213
+ ```py
214
+ model.generation_config.top_p = 0.5
215
+ ```
216
+
217
+ 特别的,可以用如下方式关闭 top-p sampling,改用 greedy sampling,效果上相当于 top_p=0 或 temperature=0:
218
+ ```py
219
+ model.generation_config.do_sample = False # greedy decoding
220
+ ```
221
+
222
+ 此外,我们在 `model.chat()` 接口也提供了调整 top_p 等参数的接口。
223
+
224
+ **有解析Action、Action Input的参考代码吗?**
225
+
226
+ 有的,可以参考:
227
+ ```py
228
+ def parse_latest_plugin_call(text: str) -> Tuple[str, str]:
229
+ i = text.rfind('\nAction:')
230
+ j = text.rfind('\nAction Input:')
231
+ k = text.rfind('\nObservation:')
232
+ if 0 <= i < j: # If the text has `Action` and `Action input`,
233
+ if k < j: # but does not contain `Observation`,
234
+ # then it is likely that `Observation` is ommited by the LLM,
235
+ # because the output text may have discarded the stop word.
236
+ text = text.rstrip() + '\nObservation:' # Add it back.
237
+ k = text.rfind('\nObservation:')
238
+ if 0 <= i < j < k:
239
+ plugin_name = text[i + len('\nAction:'):j].strip()
240
+ plugin_args = text[j + len('\nAction Input:'):k].strip()
241
+ return plugin_name, plugin_args
242
+ return '', ''
243
+ ```
244
+
245
+ 此外,如果输出的 Action Input 内容是一段表示 JSON 对象的文本,我们建议使用 `json5` 包的 `json5.loads(...)` 方法加载。
modeling_qwen.py CHANGED
@@ -958,12 +958,14 @@ class QWenLMHeadModel(QWenPreTrainedModel):
958
  history: Optional[HistoryType],
959
  system: str = "You are a helpful assistant.",
960
  append_history: bool = True,
961
- stream: Optional[bool] = False
 
 
962
  ) -> Tuple[str, HistoryType]:
963
-
964
-
965
  if history is None:
966
  history = []
 
 
967
 
968
  raw_text, context_tokens = make_context(
969
  tokenizer,
@@ -974,9 +976,9 @@ class QWenLMHeadModel(QWenPreTrainedModel):
974
  chat_format=self.generation_config.chat_format,
975
  )
976
 
977
- stop_words_ids = get_stop_words_ids(
978
  self.generation_config.chat_format, tokenizer
979
- )
980
  input_ids = torch.tensor([context_tokens]).to(self.device)
981
  if stream:
982
  assert self.generation_config.chat_format == 'chatml'
@@ -986,7 +988,8 @@ class QWenLMHeadModel(QWenPreTrainedModel):
986
  stream_config = StreamGenerationConfig(**self.generation_config.to_dict(), do_stream=True)
987
  def stream_generator():
988
  outputs = []
989
- for token in self.generate(input_ids, return_dict_in_generate=False, generation_config=stream_config):
 
990
  outputs.append(token.item())
991
  if outputs[-1] in (tokenizer.im_end_id, tokenizer.im_start_id):
992
  break
@@ -998,6 +1001,7 @@ class QWenLMHeadModel(QWenPreTrainedModel):
998
  input_ids,
999
  stop_words_ids = stop_words_ids,
1000
  return_dict_in_generate = False,
 
1001
  )
1002
 
1003
  response = decode_tokens(
 
958
  history: Optional[HistoryType],
959
  system: str = "You are a helpful assistant.",
960
  append_history: bool = True,
961
+ stream: Optional[bool] = False,
962
+ stop_words_ids: Optional[List[List[int]]] = None,
963
+ **kwargs,
964
  ) -> Tuple[str, HistoryType]:
 
 
965
  if history is None:
966
  history = []
967
+ if stop_words_ids is None:
968
+ stop_words_ids = []
969
 
970
  raw_text, context_tokens = make_context(
971
  tokenizer,
 
976
  chat_format=self.generation_config.chat_format,
977
  )
978
 
979
+ stop_words_ids.extend(get_stop_words_ids(
980
  self.generation_config.chat_format, tokenizer
981
+ ))
982
  input_ids = torch.tensor([context_tokens]).to(self.device)
983
  if stream:
984
  assert self.generation_config.chat_format == 'chatml'
 
988
  stream_config = StreamGenerationConfig(**self.generation_config.to_dict(), do_stream=True)
989
  def stream_generator():
990
  outputs = []
991
+ for token in self.generate(
992
+ input_ids, return_dict_in_generate=False, generation_config=stream_config, **kwargs):
993
  outputs.append(token.item())
994
  if outputs[-1] in (tokenizer.im_end_id, tokenizer.im_start_id):
995
  break
 
1001
  input_ids,
1002
  stop_words_ids = stop_words_ids,
1003
  return_dict_in_generate = False,
1004
+ **kwargs,
1005
  )
1006
 
1007
  response = decode_tokens(