xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-24 15:46:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

[v0][Core] Use xgrammar shared context to avoid copy overhead for offline engine (#13837)

Browse Source 
Signed-off-by: Seth Kimmel <seth.kimmel3@gmail.com>
This commit is contained in:
Seth Kimmel 2025-02-25 22:58:24 -08:00 committed by GitHub
parent 1d35662e6d
commit e206b54331
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 23 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
vllm/model_executor/guided_decoding/xgrammar_decoding.py
View File

@ -3,7 +3,6 @@
# noqa: UP007
from __future__ import annotations
import copy
import json
import re
from dataclasses import dataclass, field
@ -348,5 +347,26 @@ class XGrammarLogitsProcessor:
return scores
def clone(self) -> XGrammarLogitsProcessor:
"""Deepcopy due to per-sequence state in the matchers"""
return copy.deepcopy(self)
"""Create a new instance with shared compiled grammar
but separate state"""
new_processor = XGrammarLogitsProcessor(self.config)
# Share the compiled grammar context (immutable after compilation)
new_processor.ctx = self.ctx
# Create fresh matchers for the new sequence
if self.ctx is not None:
new_processor.matchers = [
xgr.GrammarMatcher(self.ctx) for _ in range(self.batch_size)
]
# Create a new token bitmask with the same size
if hasattr(self, 'token_bitmask') and self.token_bitmask is not None:
new_processor.token_bitmask = self.token_bitmask
# Copy simple attributes
new_processor.batch_size = self.batch_size
# Reset prefilled state for new sequence
new_processor.prefilled = False
return new_processor