[Perf][Kernel] Integrate DeepSelect TopK for the DSA sparse indexer (#56464)

Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Kimi Code <noreply@moonshot.ai>
Co-authored-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Jiangyun Zhu
2026-09-13 00:44:37 -07:00
committed by GitHub
co-authored by Kimi Code Roger Wang
parent 1cfd972816
commit fa008bdccf
8 changed files with 829 additions and 52 deletions
+16 -1
View File
@@ -76,7 +76,12 @@ from vllm.config.cache import (
PrefixCachingHashAlgo,
)
from vllm.config.device import Device
from vllm.config.kernel import IrOpPriorityConfig, LinearBackend, MoEBackend
from vllm.config.kernel import (
IrOpPriorityConfig,
LinearBackend,
MoEBackend,
SparseIndexerTopkBackend,
)
from vllm.config.load import SafetensorsLoadStrategy
from vllm.config.lora import MaxLoRARanks
from vllm.config.mamba import MambaBackendEnum, MambaSSUAlgorithm
@@ -505,6 +510,9 @@ class EngineArgs:
enable_ep_weight_filter: bool = ParallelConfig.enable_ep_weight_filter
moe_backend: MoEBackend = KernelConfig.moe_backend
linear_backend: LinearBackend = KernelConfig.linear_backend
sparse_indexer_topk_backend: SparseIndexerTopkBackend = (
KernelConfig.sparse_indexer_topk_backend
)
all2all_backend: All2AllBackend = ParallelConfig.all2all_backend
enable_elastic_ep: bool = ParallelConfig.enable_elastic_ep
elastic_ep_max_dp_size: int = ParallelConfig.elastic_ep_max_dp_size
@@ -1670,6 +1678,11 @@ class EngineArgs:
linear_backend_kwargs = kernel_kwargs["linear_backend"]
linear_backend_kwargs["type"] = lambda s: s.lower().replace("-", "_")
kernel_group.add_argument("--linear-backend", **linear_backend_kwargs)
sparse_indexer_topk_kwargs = kernel_kwargs["sparse_indexer_topk_backend"]
sparse_indexer_topk_kwargs["type"] = lambda s: s.lower().replace("-", "_")
kernel_group.add_argument(
"--sparse-indexer-topk-backend", **sparse_indexer_topk_kwargs
)
# vLLM arguments
vllm_kwargs = get_kwargs(VllmConfig)
@@ -2548,6 +2561,8 @@ class EngineArgs:
kernel_config.moe_backend = self.moe_backend
if self.linear_backend != "auto":
kernel_config.linear_backend = self.linear_backend
if self.sparse_indexer_topk_backend != "auto":
kernel_config.sparse_indexer_topk_backend = self.sparse_indexer_topk_backend
# Transfer top-level ir_op_priority into KernelConfig.ir_op_priority
for op_name, op_priority in asdict(self.ir_op_priority).items():