[Bugfix] fix automatic prefix args and add log info (#3608)

vllm/core/block_manager.py

@@ -9,6 +9,9 @@ from vllm.block import BlockTable, PhysicalTokenBlock
 from vllm.sequence import Sequence, SequenceGroup, SequenceStatus
 from vllm.utils import Device
 from vllm.core.evictor import Evictor, EvictionPolicy, make_evictor
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
 
 
 class BlockAllocatorBase(ABC):
@@ -241,11 +244,13 @@ class BlockSpaceManager:
         self.watermark_blocks = int(watermark * num_gpu_blocks)
 
         if self.enable_caching:
+            logger.info("enable automatic prefix caching")
             self.gpu_allocator = CachedBlockAllocator(Device.GPU, block_size,
                                                       num_gpu_blocks)
             self.cpu_allocator = CachedBlockAllocator(Device.CPU, block_size,
                                                       num_cpu_blocks)
         else:
+            logger.info("disable automatic prefix caching")
             self.gpu_allocator = UncachedBlockAllocator(
                 Device.GPU, block_size, num_gpu_blocks)
             self.cpu_allocator = UncachedBlockAllocator(

vllm/engine/arg_utils.py

@@ -337,7 +337,8 @@ class EngineArgs:
         cache_config = CacheConfig(self.block_size,
                                    self.gpu_memory_utilization,
                                    self.swap_space, self.kv_cache_dtype,
-                                   model_config.get_sliding_window())
+                                   model_config.get_sliding_window(),
+                                   self.enable_prefix_caching)
         parallel_config = ParallelConfig(
             self.pipeline_parallel_size, self.tensor_parallel_size,
             self.worker_use_ray, self.max_parallel_loading_workers,