[Core] Support dynamically loading Lora adapter from HuggingFace (#6234)

Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>
This commit is contained in:
Jiaxin Shan
2024-07-22 15:42:40 -07:00
committed by GitHub
co-authored by Antoni Baum
parent 69d5ae38dc
commit 42c7f66a38
11 changed files with 201 additions and 18 deletions
+4 -3
View File
@@ -13,6 +13,7 @@ from vllm.logger import init_logger
from vllm.lora.models import (LoRAModel, LoRAModelManager,
LRUCacheLoRAModelManager, create_lora_manager)
from vllm.lora.request import LoRARequest
from vllm.lora.utils import get_adapter_absolute_path
logger = init_logger(__name__)
@@ -89,8 +90,9 @@ class WorkerLoRAManager(AbstractWorkerManager):
packed_modules_mapping[module])
else:
expected_lora_modules.append(module)
lora_path = get_adapter_absolute_path(lora_request.lora_path)
lora = self._lora_model_cls.from_local_checkpoint(
lora_request.lora_local_path,
lora_path,
expected_lora_modules,
max_position_embeddings=self.max_position_embeddings,
lora_model_id=lora_request.lora_int_id,
@@ -102,8 +104,7 @@ class WorkerLoRAManager(AbstractWorkerManager):
embedding_padding_modules=self.embedding_padding_modules,
)
except Exception as e:
raise RuntimeError(
f"Loading lora {lora_request.lora_local_path} failed") from e
raise RuntimeError(f"Loading lora {lora_path} failed") from e
if lora.rank > self.lora_config.max_lora_rank:
raise ValueError(
f"LoRA rank {lora.rank} is greater than max_lora_rank "