Truncation control for embedding models (#14776)

Signed-off-by: Gabriel Marinho <gmarinho@ibm.com>
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Co-authored-by: Max de Bayser <mbayser@br.ibm.com>
This commit is contained in:
Gabriel Marinho
2025-04-30 09:24:57 +08:00
committed by GitHub
co-authored by Max de Bayser
parent 4055130a85
commit 1c2bc7ead0
21 changed files with 333 additions and 71 deletions
+13 -3
View File
@@ -55,6 +55,8 @@ def encode_tokens(
tokenizer: AnyTokenizer,
text: str,
*,
truncation: Optional[bool] = None,
max_length: Optional[int] = None,
add_special_tokens: Optional[bool] = None,
) -> list[int]:
"""
@@ -64,10 +66,18 @@ def encode_tokens(
:code:`add_special_tokens=None` means to use the backend's default
settings.
"""
if add_special_tokens is not None:
return tokenizer.encode(text, add_special_tokens=add_special_tokens)
return tokenizer.encode(text)
kw_args: dict[str, Any] = {}
if max_length is not None:
kw_args["max_length"] = max_length
if truncation is not None:
kw_args["truncation"] = truncation
if add_special_tokens is not None:
kw_args["add_special_tokens"] = add_special_tokens
return tokenizer.encode(text, **kw_args)
def get_cached_tokenizer(tokenizer: AnyTokenizer) -> AnyTokenizer:
@@ -94,6 +94,8 @@ class TokenizerBase(ABC):
@abstractmethod
def encode(self,
text: str,
truncation: Optional[bool] = None,
max_length: Optional[int] = None,
add_special_tokens: Optional[bool] = None) -> list[int]:
raise NotImplementedError()
@@ -45,11 +45,16 @@ class TokenizerGroup:
def encode(self,
prompt: str,
max_length: Optional[int] = None,
truncation: Optional[bool] = None,
lora_request: Optional[LoRARequest] = None,
add_special_tokens: Optional[bool] = None) -> List[int]:
tokenizer = self.get_lora_tokenizer(lora_request)
ret = encode_tokens(tokenizer,
prompt,
max_length=max_length,
truncation=truncation,
add_special_tokens=add_special_tokens)
self._raise_if_input_too_long(ret, lora_request)
return ret
@@ -57,11 +62,15 @@ class TokenizerGroup:
async def encode_async(
self,
prompt: str,
max_length: Optional[int] = None,
truncation: Optional[bool] = None,
lora_request: Optional[LoRARequest] = None,
add_special_tokens: Optional[bool] = None) -> List[int]:
tokenizer = await self.get_lora_tokenizer_async(lora_request)
ret = encode_tokens(tokenizer,
prompt,
max_length=max_length,
truncation=truncation,
add_special_tokens=add_special_tokens)
self._raise_if_input_too_long(ret, lora_request)
return ret
@@ -359,6 +359,8 @@ class MistralTokenizer(TokenizerBase):
def encode(self,
text: str,
truncation: Optional[bool] = None,
max_length: Optional[int] = None,
add_special_tokens: Optional[bool] = None) -> List[int]:
# `encode` should only be used for prompt completion
# it should never be used for chat_completion.