A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 10:45 UTC
[BugFix][Core] Fix BlockManagerV2 when Encoder Input is None (#9103)
sroy745 committed
Oct 6, 2024 at 20:52 UTC
c8f26bb63694adb4202ab275efb0759c13edcaa8
3 files changed
+3
-8
vllm/core/block/block_table.py
-2
@@ -220,7 +220,6 @@ class BlockTable:
220
occupied by each block. After freeing all the blocks, the `_blocks` list
221
is set to `None`.
222
"""
223
- assert self._is_allocated
223
for block in self.blocks:
224
self._allocator.free(block)
225
self._blocks.reset()
@@ -239,7 +238,6 @@ class BlockTable:
238
List[int]: A list of physical block indices for the blocks in the
239
BlockTable.
240
"""
242
- assert self._is_allocated
241
return self._blocks.ids()
242
243
def get_unseen_token_ids(self, sequence_token_ids: List[int]) -> List[int]:
vllm/core/block_manager_v2.py
+3
-1
@@ -151,7 +151,9 @@ class BlockSpaceManagerV2(BlockSpaceManager):
151
block_allocator=self.block_allocator,
152
max_block_sliding_window=self.max_block_sliding_window,
153
)
154
- block_table.allocate(seq.get_token_ids())
154
+ if seq.get_token_ids():
155
+ # Add blocks to the block table only if the sequence is non empty.
156
+ block_table.allocate(seq.get_token_ids())
157
158
return block_table
159
vllm/engine/arg_utils.py
-5
@@ -903,11 +903,6 @@ class EngineArgs:
903
"--enable-prefix-caching is currently not "
904
"supported for multimodal models and has been disabled.")
905
self.enable_prefix_caching = False
906
- if model_config.is_encoder_decoder_model:
907
- logger.warning(
908
- "Block Manager v2 does not support encoder-decoder models"
909
- " currently. Using Block Manager v1 as fallback.")
910
- self.use_v2_block_manager = False
906
907
cache_config = CacheConfig(
908
block_size=self.block_size if self.device != "neuron" else