Spaces:
Sleeping
Sleeping
File size: 6,664 Bytes
a5ab2ca |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 |
import asyncio
from typing import List, Union
from lagent.llms.base_llm import AsyncBaseLLM, BaseLLM
from lagent.utils.util import filter_suffix
def asdict_completion(output):
return {
key: getattr(output, key)
for key in [
'text', 'token_ids', 'cumulative_logprob', 'logprobs',
'finish_reason', 'stop_reason'
]
}
class VllmModel(BaseLLM):
"""
A wrapper of vLLM model.
Args:
path (str): The path to the model.
It could be one of the following options:
- i) A local directory path of a huggingface model.
- ii) The model_id of a model hosted inside a model repo
on huggingface.co, such as "internlm/internlm-chat-7b",
"Qwen/Qwen-7B-Chat ", "baichuan-inc/Baichuan2-7B-Chat"
and so on.
tp (int): tensor parallel
vllm_cfg (dict): Other kwargs for vllm model initialization.
"""
def __init__(self, path: str, tp: int = 1, vllm_cfg=dict(), **kwargs):
super().__init__(path=path, **kwargs)
from vllm import LLM
self.model = LLM(
model=self.path,
trust_remote_code=True,
tensor_parallel_size=tp,
**vllm_cfg)
def generate(self,
inputs: Union[str, List[str]],
do_preprocess: bool = None,
skip_special_tokens: bool = False,
return_dict: bool = False,
**kwargs):
"""Return the chat completions in non-stream mode.
Args:
inputs (Union[str, List[str]]): input texts to be completed.
do_preprocess (bool): whether pre-process the messages. Default to
True, which means chat_template will be applied.
skip_special_tokens (bool): Whether or not to remove special tokens
in the decoding. Default to be False.
Returns:
(a list of/batched) text/chat completion
"""
from vllm import SamplingParams
batched = True
if isinstance(inputs, str):
inputs = [inputs]
batched = False
prompt = inputs
gen_params = self.update_gen_params(**kwargs)
max_new_tokens = gen_params.pop('max_new_tokens')
stop_words = gen_params.pop('stop_words')
sampling_config = SamplingParams(
skip_special_tokens=skip_special_tokens,
max_tokens=max_new_tokens,
stop=stop_words,
**gen_params)
response = self.model.generate(prompt, sampling_params=sampling_config)
texts = [resp.outputs[0].text for resp in response]
# remove stop_words
texts = filter_suffix(texts, self.gen_params.get('stop_words'))
for resp, text in zip(response, texts):
resp.outputs[0].text = text
if batched:
return [asdict_completion(resp.outputs[0])
for resp in response] if return_dict else texts
return asdict_completion(
response[0].outputs[0]) if return_dict else texts[0]
class AsyncVllmModel(AsyncBaseLLM):
"""
A asynchronous wrapper of vLLM model.
Args:
path (str): The path to the model.
It could be one of the following options:
- i) A local directory path of a huggingface model.
- ii) The model_id of a model hosted inside a model repo
on huggingface.co, such as "internlm/internlm-chat-7b",
"Qwen/Qwen-7B-Chat ", "baichuan-inc/Baichuan2-7B-Chat"
and so on.
tp (int): tensor parallel
vllm_cfg (dict): Other kwargs for vllm model initialization.
"""
def __init__(self, path: str, tp: int = 1, vllm_cfg=dict(), **kwargs):
super().__init__(path=path, **kwargs)
from vllm import AsyncEngineArgs, AsyncLLMEngine
engine_args = AsyncEngineArgs(
model=self.path,
trust_remote_code=True,
tensor_parallel_size=tp,
**vllm_cfg)
self.model = AsyncLLMEngine.from_engine_args(engine_args)
async def generate(self,
inputs: Union[str, List[str]],
session_ids: Union[int, List[int]] = None,
do_preprocess: bool = None,
skip_special_tokens: bool = False,
return_dict: bool = False,
**kwargs):
"""Return the chat completions in non-stream mode.
Args:
inputs (Union[str, List[str]]): input texts to be completed.
do_preprocess (bool): whether pre-process the messages. Default to
True, which means chat_template will be applied.
skip_special_tokens (bool): Whether or not to remove special tokens
in the decoding. Default to be False.
Returns:
(a list of/batched) text/chat completion
"""
from vllm import SamplingParams
batched = True
if isinstance(inputs, str):
inputs = [inputs]
batched = False
if session_ids is None:
session_ids = list(range(len(inputs)))
elif isinstance(session_ids, (int, str)):
session_ids = [session_ids]
assert len(inputs) == len(session_ids)
prompt = inputs
gen_params = self.update_gen_params(**kwargs)
max_new_tokens = gen_params.pop('max_new_tokens')
stop_words = gen_params.pop('stop_words')
sampling_config = SamplingParams(
skip_special_tokens=skip_special_tokens,
max_tokens=max_new_tokens,
stop=stop_words,
**gen_params)
async def _inner_generate(uid, text):
resp, generator = '', self.model.generate(
text, sampling_params=sampling_config, request_id=uid)
async for out in generator:
resp = out.outputs[0]
return resp
response = await asyncio.gather(*[
_inner_generate(sid, inp) for sid, inp in zip(session_ids, prompt)
])
texts = [resp.text for resp in response]
# remove stop_words
texts = filter_suffix(texts, self.gen_params.get('stop_words'))
for resp, text in zip(response, texts):
resp.text = text
if batched:
return [asdict_completion(resp)
for resp in response] if return_dict else texts
return asdict_completion(response[0]) if return_dict else texts[0]
|