biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Core] Use sha256 bytes instead of BlockHash to reduce GC overhead (#23673)

Browse Source 
Signed-off-by: linzebing <linzebing1995@gmail.com>
This commit is contained in:
Zebing Lin
2025-09-09 00:34:37 -04:00
committed by GitHub
parent bba1042c6f
commit 82dfb12e52
15 changed files with 298 additions and 283 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
vllm/engine/arg_utils.py
View File

@@ -1592,20 +1592,12 @@ class EngineArgs:
"in low performance due to small KV cache size. Consider "
"setting --max-model-len to a smaller value.", max_model_len)
# if using prefix caching, we must set a hash algo
if self.enable_prefix_caching:
# Disable prefix caching for multimodal models for VLLM_V0.
if model_config.is_multimodal_model:
logger.warning(
"--enable-prefix-caching is not supported for multimodal "
"models in V0 and has been disabled.")
self.enable_prefix_caching = False
# VLLM_V0 only supports builtin hash algo for prefix caching.
if self.prefix_caching_hash_algo == "sha256":
raise ValueError(
"sha256 is not supported for prefix caching in V0 engine. "
"Please use 'builtin'.")
# Disable prefix caching for multimodal models for VLLM_V0.
if self.enable_prefix_caching and model_config.is_multimodal_model:
logger.warning(
"--enable-prefix-caching is not supported for multimodal "
"models in V0 and has been disabled.")
self.enable_prefix_caching = False
# Set max_num_seqs to 256 for VLLM_V0.
if self.max_num_seqs is None: