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] use blockmanagerv1 for encoder-decoder (#9084)
Co-authored-by: Roger Wang <ywang@roblox.com>
Chen Zhang committed
Oct 4, 2024 at 21:45 UTC
dac914b0d6bc36de4eb4bf70a9d20954560893ea
1 file changed
+5
vllm/engine/arg_utils.py
+5
@@ -903,6 +903,11 @@ 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
911
912
cache_config = CacheConfig(
913
block_size=self.block_size if self.device != "neuron" else