github/DB-GPT
1
0
Fork 0
mirror of https://github.com/csunny/DB-GPT.git synced 2025-08-28 12:51:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: fix vllm tokenizer count error (#2555)

Browse Source 
Co-authored-by: Fangyin Cheng <staneyffer@gmail.com>
This commit is contained in:
magic.chen 2025-03-31 08:31:04 +08:00 committed by GitHub
parent c68332be4c
commit 0bdff7a0f2
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 18 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
packages/dbgpt-core/src/dbgpt/model/cluster/worker/default_worker.py
View File

@ -19,6 +19,8 @@ from dbgpt.model.adapter.base import LLMModelAdapter
from dbgpt.model.adapter.loader import ModelLoader
from dbgpt.model.adapter.model_adapter import get_llm_model_adapter
from dbgpt.model.cluster.worker_base import ModelWorker
from dbgpt.model.proxy.base import TiktokenProxyTokenizer
from dbgpt.util.executor_utils import blocking_func_to_async_no_executor
from dbgpt.util.model_utils import _clear_model_cache, _get_current_cuda_memory
from dbgpt.util.parameter_utils import _get_dict_from_obj
from dbgpt.util.system_utils import get_system_info
@ -43,6 +45,8 @@ class DefaultModelWorker(ModelWorker):
self._support_generate_func = False
self.context_len = 4096
self._device = get_device()
# Use tiktoken to count token if model doesn't support
self._tiktoken = TiktokenProxyTokenizer()
def load_worker(
self, model_name: str, deploy_model_params: BaseDeployModelParameters, **kwargs
@ -241,18 +245,20 @@ class DefaultModelWorker(ModelWorker):
return output
def count_token(self, prompt: str) -> int:
return _try_to_count_token(prompt, self.tokenizer, self.model)
return _try_to_count_token(prompt, self.tokenizer, self.model, self._tiktoken)
async def async_count_token(self, prompt: str) -> int:
# TODO if we deploy the model by vllm, it can't work, we should run
# transformer _try_to_count_token to async
from dbgpt.model.proxy.llms.proxy_model import ProxyModel
if isinstance(self.model, ProxyModel) and self.model.proxy_llm_client:
return await self.model.proxy_llm_client.count_token(
self.model.proxy_llm_client.default_model, prompt
)
raise NotImplementedError
cnt = await blocking_func_to_async_no_executor(
_try_to_count_token, prompt, self.tokenizer, self.model, self._tiktoken
)
return cnt
def get_model_metadata(self, params: Dict) -> ModelMetadata:
ext_metadata = ModelExtraMedata(
@ -594,7 +600,9 @@ def _new_metrics_from_model_output(
return metrics
def _try_to_count_token(prompt: str, tokenizer, model) -> int:
def _try_to_count_token(
prompt: str, tokenizer, model, tiktoken: TiktokenProxyTokenizer
) -> int:
"""Try to count token of prompt
Args:
@ -612,11 +620,11 @@ def _try_to_count_token(prompt: str, tokenizer, model) -> int:
if isinstance(model, ProxyModel):
return model.count_token(prompt)
# Only support huggingface model now
return len(tokenizer(prompt).input_ids[0])
except Exception as e:
logger.warning(f"Count token error, detail: {e}, return -1")
return -1
# Only support huggingface and vllm model now
return len(tokenizer([prompt]).input_ids[0])
except Exception as _e:
logger.warning("Failed to count token, try tiktoken")
return tiktoken.count_token("cl100k_base", [prompt])[0]
def _try_import_torch():