Introduce Stable LoRA ID System for Overlapped Updates and Prefix Caching (#8261)

This commit is contained in:
Lifu Huang
2025-07-23 00:32:16 -07:00
committed by GitHub
parent e885bfdc6a
commit 8abd3e77fe
11 changed files with 400 additions and 261 deletions

View File

@@ -22,6 +22,7 @@ from dataclasses import dataclass, field
from enum import Enum
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
from sglang.srt.lora.lora_registry import LoRARef
from sglang.srt.managers.schedule_batch import BaseFinishReason
from sglang.srt.multimodal.mm_utils import has_valid_data
from sglang.srt.sampling.sampling_params import SamplingParams
@@ -1067,19 +1068,36 @@ class LoadLoRAAdapterReqInput:
lora_name: str
# The path of loading.
lora_path: str
# The unique identifier for the LoRA adapter, which automatically generated in the `TokenizerManager`.
lora_id: Optional[str] = None
def to_ref(self) -> LoRARef:
return LoRARef(
lora_id=self.lora_id,
lora_name=self.lora_name,
lora_path=self.lora_path,
)
@dataclass
class UnloadLoRAAdapterReqInput:
# The name of lora module to unload.
lora_name: str
# The unique identifier for the LoRA adapter, which automatically generated in the `TokenizerManager`.
lora_id: Optional[str] = None
def to_ref(self) -> LoRARef:
return LoRARef(
lora_id=self.lora_id,
lora_name=self.lora_name,
)
@dataclass
class LoRAUpdateResult:
success: bool
error_message: Optional[str] = None
loaded_adapters: Dict[str, str] = field(default_factory=dict)
loaded_adapters: Dict[str, LoRARef] = field(default_factory=dict)
LoadLoRAAdapterReqOutput = UnloadLoRAAdapterReqOutput = LoRAUpdateResult