[Platform] Support platform-specific MLA prefill backend selection (#56959)

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
wangxiyuan
2026-10-08 15:25:21 +00:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent fd928e5f3c
commit 072d49bb6e
11 changed files with 469 additions and 283 deletions
+20
View File
@@ -18,6 +18,7 @@ from vllm.utils.cpu_resource_utils import (
get_visible_memory_node,
)
from vllm.utils.mem_constants import GiB_bytes
from vllm.v1.attention.backends.mla.prefill.registry import MLAPrefillBackendEnum
from vllm.v1.attention.backends.registry import AttentionBackendEnum
from .interface import CpuArchEnum, Platform, PlatformEnum
@@ -26,6 +27,10 @@ logger = init_logger(__name__)
if TYPE_CHECKING:
from vllm.config import VllmConfig
from vllm.v1.attention.backends.mla.prefill.base import (
MLAPrefillBackend,
)
from vllm.v1.attention.backends.mla.prefill.selector import MLAPrefillSelectorConfig
from vllm.v1.attention.selector import AttentionSelectorConfig
else:
VllmConfig = None
@@ -173,6 +178,21 @@ class CpuPlatform(Platform):
logger.info("Cannot use %s backend on CPU.", selected_backend)
return AttentionBackendEnum.CPU_ATTN.get_path()
@classmethod
def get_mla_prefill_backend_cls(
cls,
mla_selector_config: "MLAPrefillSelectorConfig",
) -> "type[MLAPrefillBackend]":
"""Get the MLA prefill backend class of a device."""
for backend_enum in (MLAPrefillBackendEnum.ZEN_CPU, MLAPrefillBackendEnum.CPU):
try:
cpu_backend_cls = backend_enum.get_class()
except ImportError:
continue
if cpu_backend_cls.is_available():
return cpu_backend_cls
raise ValueError("No valid CPU MLA prefill backend found.")
@classmethod
def get_device_total_memory(cls, device_id: int = 0) -> int:
meminfo = get_memory_node_info(device_id)