Extend a LoRARequest to support model validation

Extend a LoRARequest with a validate() method to enable validation of a
LoRA adapter when it is loaded.

Add a test case.

Signed-off-by: Stefan Berger <[email protected]>

Author: stefanberger

tests/lora/test_lora_manager.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import os
+from typing import Optional
 
 import pytest
 import torch
@@ -26,6 +27,11 @@
 from vllm.lora.request import LoRARequest
 from vllm.lora.worker_manager import LRUCacheWorkerLoRAManager, WorkerLoRAManager
 from vllm.platforms import current_platform
+from vllm.validation.plugins import (
+    ModelType,
+    ModelValidationPlugin,
+    ModelValidationPluginRegistry,
+)
 
 from .utils import create_peft_lora
 
@@ -714,3 +720,36 @@ def test_packed_loras(dist_init, dummy_model_gate_up, device):
     torch.testing.assert_close(
         packed_lora1.lora_b[1], model_lora_clone1.get_lora("up_proj").lora_b
     )
+
+
+class MyModelValidator(ModelValidationPlugin):
+    def model_validation_needed(self, model_type: ModelType, model_path: str) -> bool:
+        return True
+
+    def validate_model(
+        self, model_type: ModelType, model_path: str, model: Optional[str] = None
+    ) -> None:
+        raise BaseException("Model did not validate")
+
+
+def test_worker_adapter_manager_security_policy(dist_init, dummy_model_gate_up):
+    my_model_validator = MyModelValidator()
+    ModelValidationPluginRegistry.register_plugin("test", my_model_validator)
+
+    lora_config = LoRAConfig(
+        max_lora_rank=8, max_cpu_loras=4, max_loras=4, lora_dtype=DEFAULT_DTYPE
+    )
+
+    model_config = ModelConfig(max_model_len=16)
+    vllm_config = VllmConfig(model_config=model_config, lora_config=lora_config)
+
+    worker_adapter_manager = WorkerLoRAManager(
+        vllm_config, "cpu", EMBEDDING_MODULES, EMBEDDING_PADDING_MODULES
+    )
+    worker_adapter_manager.create_lora_manager(dummy_model_gate_up)
+
+    mapping = LoRAMapping([], [])
+    with pytest.raises(BaseException, match="Model did not validate"):
+        worker_adapter_manager.set_active_adapters(
+            [LoRARequest("1", 1, "/not/used")], mapping
+        )

vllm/lora/request.py

@@ -6,6 +6,8 @@
 
 import msgspec
 
+from vllm.validation.plugins import ModelType, ModelValidationPluginRegistry
+
 
 class LoRARequest(
     msgspec.Struct,
@@ -99,3 +101,9 @@ def __hash__(self) -> int:
         identified by their names across engines.
         """
         return hash(self.lora_name)
+
+    def validate(self) -> None:
+        """Validate the LoRA adapter given its path."""
+        ModelValidationPluginRegistry.validate_model(
+            ModelType.MODEL_TYPE_LORA, self.lora_path
+        )

vllm/lora/worker_manager.py

@@ -85,6 +85,7 @@ def create_lora_manager(
         return lora_manager.model
 
     def _load_adapter(self, lora_request: LoRARequest) -> LoRAModel:
+        lora_request.validate()
         try:
             supported_lora_modules = self._adapter_manager.supported_lora_modules
             packed_modules_mapping = self._adapter_manager.packed_modules_mapping