Implement LRU eviction policy for LoRA adapters (#11041)

This commit is contained in:
Chenxi Li
2025-10-13 20:18:25 -07:00
committed by GitHub
parent 88a6f9dab5
commit 28f80b1244
9 changed files with 399 additions and 15 deletions

View File

@@ -519,6 +519,7 @@ class SRTRunner:
lora_target_modules: Optional[List[str]] = None,
enable_lora: Optional[bool] = None,
max_loaded_loras: Optional[int] = None,
lora_eviction_policy: str = "lru",
):
self.model_type = model_type
self.is_generation = model_type == "generation"
@@ -565,6 +566,7 @@ class SRTRunner:
lora_target_modules=lora_target_modules,
enable_lora=enable_lora,
max_loaded_loras=max_loaded_loras,
lora_eviction_policy=lora_eviction_policy,
**spec_kwargs,
)