Reapply [Attention] Refactor check_and_update_config (#35122)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Matthew Bonanni
2026-03-09 07:17:14 -07:00
committed by GitHub
parent 5578f2a4d3
commit 77a73458e3
32 changed files with 311 additions and 279 deletions
+50
View File
@@ -420,6 +420,56 @@ class Platform:
"""
pass
@classmethod
def update_block_size_for_backend(cls, vllm_config: "VllmConfig") -> None:
"""
Ensure block_size is compatible with the attention backend.
"""
from vllm.config.cache import CacheConfig
cache_config = vllm_config.cache_config
if cache_config.user_specified_block_size:
# User specified --block-size; keep it.
return
model_config = vllm_config.model_config
# model_config may be None during testing.
# Skip hybrid models — their block_size is managed by
# HybridAttentionMambaModelConfig.
if model_config is None or model_config.is_hybrid:
cache_config.block_size = CacheConfig.DEFAULT_BLOCK_SIZE
return
from vllm.config.vllm import (
get_layers_from_vllm_config,
set_current_vllm_config,
)
from vllm.model_executor.layers.attention_layer_base import (
AttentionLayerBase,
)
attn_layers = get_layers_from_vllm_config(
vllm_config,
AttentionLayerBase, # type: ignore[type-abstract]
)
if not attn_layers:
cache_config.block_size = CacheConfig.DEFAULT_BLOCK_SIZE
return
first_layer = next(iter(attn_layers.values()))
backend_cls = first_layer.get_attn_backend()
with set_current_vllm_config(vllm_config):
preferred = backend_cls.get_preferred_block_size(
CacheConfig.DEFAULT_BLOCK_SIZE
)
if preferred != CacheConfig.DEFAULT_BLOCK_SIZE:
logger.info(
"Setting kv cache block size to %d for %s backend.",
preferred,
backend_cls.get_name(),
)
cache_config.block_size = preferred
@classmethod
def verify_model_arch(cls, model_arch: str) -> None:
"""