[V0][V1][Core] Add outlines integration for V1, and update V0 integration. (#15975)

Signed-off-by: Nathan Hoos <thwackyy.y@gmail.com>
This commit is contained in:
Nathan Hoos
2025-07-10 15:30:26 -04:00
committed by GitHub
parent 5e53c89a74
commit d6902ce79f
13 changed files with 804 additions and 461 deletions
@@ -12,7 +12,7 @@ from regex import escape as regex_escape
from transformers import PreTrainedTokenizerBase
from vllm.model_executor.guided_decoding.outlines_logits_processors import (
CFGLogitsProcessor, JSONLogitsProcessor, RegexLogitsProcessor)
JSONLogitsProcessor, RegexLogitsProcessor)
from vllm.reasoning import ReasoningParser
from vllm.sampling_params import GuidedDecodingParams
@@ -21,36 +21,8 @@ class GuidedDecodingMode(Enum):
JSON = "json"
REGEX = "regex"
CHOICE = "choice"
GRAMMAR = "grammar"
# https://github.com/outlines-dev/outlines/blob/main/outlines/grammars/json.lark
# the main difference is that we changed the start: value to
# start: object | array, so we are denying scalar values as the root of the
# JSON. Starting with scalars as the root seems to cause llama to generate
# without stop.
JSON_GRAMMAR = r"""
?start: object | array
?value: object
| array
| UNESCAPED_STRING
| SIGNED_NUMBER -> number
| "true" -> true
| "false" -> false
| "null" -> null
array : "[" [value ("," value)*] "]"
object : "{" [pair ("," pair)*] "}"
pair : UNESCAPED_STRING ":" value
%import common.UNESCAPED_STRING
%import common.SIGNED_NUMBER
%import common.WS
%ignore WS
"""
global_thread_pool = None # used for generating logits processor fsm
# It's not yet clear that using more provides a benefit, and it could
@@ -60,16 +32,12 @@ _MAX_THREADPOOL_WORKERS = 16
async def get_outlines_guided_decoding_logits_processor(
guided_params: GuidedDecodingParams,
tokenizer: PreTrainedTokenizerBase,
reasoner: Optional[ReasoningParser],
) -> Union[JSONLogitsProcessor, RegexLogitsProcessor, CFGLogitsProcessor,
None]:
guided_params: GuidedDecodingParams, tokenizer: PreTrainedTokenizerBase,
reasoner: Optional[ReasoningParser]
) -> Union[JSONLogitsProcessor, RegexLogitsProcessor, None]:
"""
Given an OpenAI-compatible request, check for guided decoding parameters
and get the necessary logits processor for the given guide.
We cache logit processors by (guide, tokenizer), and on cache hit
we make a shallow copy to reuse the same underlying FSM.
"""
global global_thread_pool
guide, mode = _get_guide_and_mode(guided_params)
@@ -83,7 +51,6 @@ async def get_outlines_guided_decoding_logits_processor(
global_thread_pool = concurrent.futures.ThreadPoolExecutor(
max_workers=max_workers)
loop = asyncio.get_running_loop()
return await loop.run_in_executor(global_thread_pool,
_get_logits_processor, guide, tokenizer,
mode, guided_params.whitespace_pattern,
@@ -91,16 +58,12 @@ async def get_outlines_guided_decoding_logits_processor(
def get_local_outlines_guided_decoding_logits_processor(
guided_params: GuidedDecodingParams,
tokenizer: PreTrainedTokenizerBase,
reasoner: Optional[ReasoningParser],
) -> Union[JSONLogitsProcessor, RegexLogitsProcessor, CFGLogitsProcessor,
None]:
guided_params: GuidedDecodingParams, tokenizer: PreTrainedTokenizerBase,
reasoner: Optional[ReasoningParser]
) -> Union[JSONLogitsProcessor, RegexLogitsProcessor, None]:
"""
Given an OpenAI-compatible request, check for guided decoding parameters
and get the necessary logits processor for the given guide.
We cache logit processors by (guide, tokenizer), and on cache hit
we make a shallow copy to reuse the same underlying FSM.
"""
guide, mode = _get_guide_and_mode(guided_params)
if not guide or not mode:
@@ -130,9 +93,10 @@ def _get_guide_and_mode(
choices_regex = "(" + "|".join(choices) + ")"
return choices_regex, GuidedDecodingMode.CHOICE
elif guided_params.grammar:
return guided_params.grammar, GuidedDecodingMode.GRAMMAR
elif guided_params.json_object:
return JSON_GRAMMAR, GuidedDecodingMode.GRAMMAR
raise ValueError(
"The `outlines` guided decoding backend no longer supports grammar "
"guided generation. Please use either the `xgrammar` or `guidance` "
"backend")
else:
return None, None
@@ -143,13 +107,11 @@ def _get_logits_processor(
mode: GuidedDecodingMode,
whitespace_pattern: Union[str, None],
reasoner: Optional[ReasoningParser],
) -> Union[JSONLogitsProcessor, RegexLogitsProcessor, CFGLogitsProcessor]:
) -> Union[JSONLogitsProcessor, RegexLogitsProcessor]:
if mode == GuidedDecodingMode.JSON:
return JSONLogitsProcessor(guide, tokenizer, whitespace_pattern,
reasoner)
elif mode == GuidedDecodingMode.REGEX or mode == GuidedDecodingMode.CHOICE:
return RegexLogitsProcessor(guide, tokenizer, reasoner)
elif mode == GuidedDecodingMode.GRAMMAR:
return CFGLogitsProcessor(guide, tokenizer, reasoner)
else:
raise ValueError(f"Unknown guided decoding mode {mode}")