[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
+4
View File
@@ -1015,6 +1015,7 @@ class precompiled_wheel_utils:
"vllm/_flashmla_extension_C.abi3.so",
"vllm/_flashkda_C.abi3.so",
"vllm/_sparse_flashmla_C.abi3.so",
"vllm/_deepselect_C.abi3.so",
"vllm/vllm_flash_attn/_vllm_fa2_C.abi3.so",
"vllm/vllm_flash_attn/_vllm_fa3_C.abi3.so",
"vllm/cumem_allocator.abi3.so",
@@ -1381,6 +1382,9 @@ if _is_cuda():
ext_modules.append(
CMakeExtension(name="vllm._flashmla_extension_C", optional=True)
)
# DeepSelect requires CUDA 12.9 or later (SM100a/SM103a only)
# Optional since it won't build on unsupported architectures
ext_modules.append(CMakeExtension(name="vllm._deepselect_C", optional=True))
if USE_PRECOMPILED_EXTENSIONS or (
CUDA_HOME and get_nvcc_cuda_version() >= Version("12.0")
):