[BugFix][Core] Fix BlockManagerV2 when Encoder Input is None (#9103)

vllm/core/block/block_table.py

@@ -220,7 +220,6 @@ class BlockTable:
         occupied by each block. After freeing all the blocks, the `_blocks` list
         is set to `None`.
         """
-        assert self._is_allocated
         for block in self.blocks:
             self._allocator.free(block)
         self._blocks.reset()
@@ -239,7 +238,6 @@ class BlockTable:
             List[int]: A list of physical block indices for the blocks in the
                 BlockTable.
         """
-        assert self._is_allocated
         return self._blocks.ids()
 
     def get_unseen_token_ids(self, sequence_token_ids: List[int]) -> List[int]:

vllm/core/block_manager_v2.py

@@ -151,7 +151,9 @@ class BlockSpaceManagerV2(BlockSpaceManager):
             block_allocator=self.block_allocator,
             max_block_sliding_window=self.max_block_sliding_window,
         )
-        block_table.allocate(seq.get_token_ids())
+        if seq.get_token_ids():
+            # Add blocks to the block table only if the sequence is non empty.
+            block_table.allocate(seq.get_token_ids())
 
         return block_table
 

vllm/engine/arg_utils.py

@@ -903,11 +903,6 @@ class EngineArgs:
                     "--enable-prefix-caching is currently not "
                     "supported for multimodal models and has been disabled.")
             self.enable_prefix_caching = False
-        if model_config.is_encoder_decoder_model:
-            logger.warning(
-                "Block Manager v2 does not support encoder-decoder models"
-                " currently. Using Block Manager v1 as fallback.")
-            self.use_v2_block_manager = False
 
         cache_config = CacheConfig(
             block_size=self.block_size if self.device != "neuron" else