[Feature] Add auto-detection for reasoning_config when only reasoning_parser is set (#38214)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
This commit is contained in:
@@ -5,6 +5,7 @@ from dataclasses import field
|
||||
|
||||
from vllm.config.model import ModelConfig
|
||||
from vllm.config.utils import config
|
||||
from vllm.reasoning import ReasoningParserManager
|
||||
from vllm.tokenizers import cached_tokenizer_from_config
|
||||
|
||||
|
||||
@@ -18,11 +19,11 @@ class ReasoningConfig:
|
||||
`initialize_token_ids` and are not intended to be set directly.
|
||||
"""
|
||||
|
||||
# NOTE: These parameters are temporary, the intent is to derive them
|
||||
# automatically from the reasoning parser in a future version.
|
||||
reasoning_start_str: str = "<think>"
|
||||
reasoning_parser: str = ""
|
||||
"""The name of the ReasoningParser to use for this model."""
|
||||
reasoning_start_str: str = ""
|
||||
"""String that indicates the start of reasoning."""
|
||||
reasoning_end_str: str = "</think>"
|
||||
reasoning_end_str: str = ""
|
||||
"""String that indicates the end of reasoning content."""
|
||||
|
||||
_reasoning_start_token_ids: list[int] | None = field(
|
||||
@@ -36,6 +37,16 @@ class ReasoningConfig:
|
||||
"""Private backing field for `reasoning_end_token_ids`. Set by
|
||||
`initialize_token_ids`. Not intended to be configured directly."""
|
||||
|
||||
_enabled: bool = field(default=False, init=False, repr=False)
|
||||
"""Private field indicating whether reasoning token IDs have been initialized.
|
||||
Set to True by `initialize_token_ids` once token IDs are initialized."""
|
||||
|
||||
@property
|
||||
def enabled(self) -> bool:
|
||||
"""Returns True if reasoning is enabled (i.e. if token IDs have been
|
||||
initialized), False otherwise."""
|
||||
return self._enabled
|
||||
|
||||
@property
|
||||
def reasoning_start_token_ids(self) -> list[int] | None:
|
||||
"""Token IDs derived from `reasoning_start_str`. Set automatically by
|
||||
@@ -54,15 +65,36 @@ class ReasoningConfig:
|
||||
self._reasoning_start_token_ids is not None
|
||||
and self._reasoning_end_token_ids is not None
|
||||
):
|
||||
return
|
||||
self._enabled = True
|
||||
return # Already initialized
|
||||
|
||||
tokenizer = cached_tokenizer_from_config(model_config=model_config)
|
||||
reasoning_start_str = self.reasoning_start_str
|
||||
reasoning_end_str = self.reasoning_end_str
|
||||
if self.reasoning_parser is not None and (
|
||||
not reasoning_start_str or not reasoning_end_str
|
||||
):
|
||||
parser_cls = ReasoningParserManager.get_reasoning_parser(
|
||||
self.reasoning_parser
|
||||
)
|
||||
reasoning_parser = parser_cls(tokenizer)
|
||||
start_token = reasoning_parser.reasoning_start_str
|
||||
if start_token and not reasoning_start_str:
|
||||
reasoning_start_str = start_token
|
||||
|
||||
end_token = reasoning_parser.reasoning_end_str
|
||||
if end_token and not reasoning_end_str:
|
||||
reasoning_end_str = end_token
|
||||
|
||||
if not reasoning_start_str or not reasoning_end_str:
|
||||
# If we don't have valid strings to tokenize,
|
||||
# we can't initialize the token IDs.
|
||||
return
|
||||
self._reasoning_start_token_ids = tokenizer.encode(
|
||||
self.reasoning_start_str, add_special_tokens=False
|
||||
reasoning_start_str, add_special_tokens=False
|
||||
)
|
||||
self._reasoning_end_token_ids = tokenizer.encode(
|
||||
self.reasoning_end_str, add_special_tokens=False
|
||||
reasoning_end_str, add_special_tokens=False
|
||||
)
|
||||
|
||||
if not self._reasoning_start_token_ids or not self._reasoning_end_token_ids:
|
||||
@@ -72,3 +104,4 @@ class ReasoningConfig:
|
||||
f"reasoning_end_str='{self.reasoning_end_str}'. "
|
||||
"Ensure the strings are valid tokens in the model's vocabulary."
|
||||
)
|
||||
self._enabled = True
|
||||
|
||||
@@ -1210,6 +1210,12 @@ class VllmConfig:
|
||||
|
||||
if self.reasoning_config is not None and self.model_config is not None:
|
||||
self.reasoning_config.initialize_token_ids(self.model_config)
|
||||
if not self.reasoning_config.enabled:
|
||||
logger.warning_once(
|
||||
"Auto-initialization of reasoning token IDs failed. "
|
||||
"Please check whether your reasoning parser has implemented "
|
||||
"the `reasoning_start_str` and `reasoning_end_str`."
|
||||
)
|
||||
|
||||
# Hybrid KV cache manager (HMA) runtime rules:
|
||||
# - Explicit enable (--no-disable-kv-cache-manager): error if runtime
|
||||
|
||||
Reference in New Issue
Block a user