Higher priority for user input of max_prefill_tokens & format (#540)

python/sglang/srt/constrained/__init__.py

@@ -4,7 +4,7 @@ from typing import Dict, Optional, Union
 from outlines.caching import cache as disk_cache
 from outlines.caching import disable_cache
 from outlines.fsm.guide import RegexGuide
-from outlines.fsm.regex import FSMInfo, make_deterministic_fsm, make_byte_level_fsm
+from outlines.fsm.regex import FSMInfo, make_byte_level_fsm, make_deterministic_fsm
 from outlines.models.transformers import TransformerTokenizer
 from pydantic import BaseModel
 

python/sglang/srt/constrained/fsm_cache.py

@@ -1,4 +1,5 @@
 """Cache for the compressed finite state machine."""
+
 from sglang.srt.constrained import RegexGuide, TransformerTokenizer
 from sglang.srt.constrained.base_cache import BaseCache
 

python/sglang/srt/constrained/jump_forward.py

@@ -8,11 +8,12 @@ from collections import defaultdict
 
 import interegular
 import outlines.caching
+
 from sglang.srt.constrained import (
     FSMInfo,
     disk_cache,
-    make_deterministic_fsm,
     make_byte_level_fsm,
+    make_deterministic_fsm,
 )
 from sglang.srt.constrained.base_cache import BaseCache