[CPU] Keep the hybrid block size across the engine core's config rerun (#61805)

Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Gabriel Wu
2026-10-17 12:06:03 +08:00
committed by GitHub
co-authored by Claude Opus 5.5
parent ce49c47cbb
commit dcd07967ca
2 changed files with 38 additions and 8 deletions
+16 -8
View File
@@ -289,10 +289,9 @@ class CpuPlatform(Platform):
cache_config.block_size,
)
cache_config.block_size = 256
elif not cache_config.user_specified_block_size:
cache_config.block_size = 128
if not reference_cpu_mla_enabled and cache_config.block_size % 32 != 0:
elif (
cache_config.user_specified_block_size and cache_config.block_size % 32 != 0
):
logger.warning(
"CPU backend prefers block_size is multiples of 32, "
"otherwise the performance is not optimized."
@@ -532,15 +531,24 @@ class CpuPlatform(Platform):
@classmethod
def update_block_size_for_backend(cls, vllm_config: "VllmConfig") -> None:
model_config = vllm_config.model_config
if model_config is None or not model_config.is_hybrid:
if model_config is None:
return
# reconcile attention and mamba page sizes
backend_classes = cls._find_non_ssm_backends(vllm_config)
if not backend_classes:
return
pre_block_size = vllm_config.cache_config.block_size
# Not in check_and_update_config(), which reruns with every
# VllmConfig.__post_init__ and would undo the hybrid alignment below.
cache_config = vllm_config.cache_config
if not cache_config.user_specified_block_size:
cache_config.block_size = cls._preferred_block_size_for_backends(
backend_classes, 128, vllm_config
)
if not model_config.is_hybrid:
return
# reconcile attention and mamba page sizes
pre_block_size = cache_config.block_size
cls._align_hybrid_block_size(vllm_config, backend_classes[0])
cls._check_aligned_block_size(vllm_config, backend_classes, pre_block_size)