diff --git a/.buildkite/check-torch-abi.py b/.buildkite/check-torch-abi.py index 6aa51d66bf..5083ad8dc5 100644 --- a/.buildkite/check-torch-abi.py +++ b/.buildkite/check-torch-abi.py @@ -13,7 +13,6 @@ from torch_abi_audit.report import ExtensionReport, PackageReport # Temporary allowlist of extensions not yet on the stable ABI. # Shrink and remove over time. ALLOWED_UNSTABLE_LIBRARIES: tuple[str, ...] = ( - "_flashkda_C.abi3.so", "vllm_flash_attn/_vllm_fa3_C.abi3.so", "third_party/deep_gemm/_C*.so", ) diff --git a/cmake/external_projects/flashkda.cmake b/cmake/external_projects/flashkda.cmake index 28b471a9e7..1d3d163c61 100644 --- a/cmake/external_projects/flashkda.cmake +++ b/cmake/external_projects/flashkda.cmake @@ -13,7 +13,7 @@ else() FetchContent_Declare( flashkda GIT_REPOSITORY https://github.com/vllm-project/FlashKDA.git - GIT_TAG a3e42bbbece3bb38f7c426b880315294a336e82f + GIT_TAG b5d11010ff01c1d4a683c0dde42e76cbeaa8107f GIT_PROGRESS TRUE GIT_SUBMODULES cutlass ) @@ -63,6 +63,12 @@ if(FLASH_KDA_ARCHS) USE_SABI 3 WITH_SOABI) + target_compile_definitions(_flashkda_C PRIVATE + TORCH_TARGET_VERSION=0x020B000000000000ULL) + if(VLLM_GPU_LANG STREQUAL "CUDA") + target_compile_definitions(_flashkda_C PRIVATE USE_CUDA) + endif() + target_compile_options(_flashkda_C PRIVATE $<$:-UPy_LIMITED_API --expt-relaxed-constexpr --expt-extended-lambda --use_fast_math -O3> $<$:-UPy_LIMITED_API>) diff --git a/csrc/flashkda_registration.cpp b/csrc/flashkda_registration.cpp index 9058119e11..f3da713a09 100644 --- a/csrc/flashkda_registration.cpp +++ b/csrc/flashkda_registration.cpp @@ -1,17 +1,24 @@ #include "core/registration.h" #include "flash_kda.h" -TORCH_LIBRARY(_flashkda_C, m) { - m.def("get_workspace_size(int T_total, int H, int N=1) -> int", - &get_workspace_size); +#include + +STABLE_TORCH_LIBRARY(_flashkda_C, m) { + m.def("get_workspace_size(int T_total, int H, int N=1) -> int"); m.def( "fwd(Tensor q, Tensor k, Tensor v, Tensor g, Tensor beta, float scale, " - "Tensor(a!) out, Tensor workspace, Tensor A_log, Tensor dt_bias, " + "Tensor(a!) out, Tensor(c!) workspace, Tensor A_log, Tensor dt_bias, " "float lower_bound, " "Tensor? initial_state=None, Tensor(b!)? final_state=None, " "Tensor? cu_seqlens=None) -> ()"); } -TORCH_LIBRARY_IMPL(_flashkda_C, CUDA, m) { m.impl("fwd", &fwd); } +STABLE_TORCH_LIBRARY_IMPL(_flashkda_C, CompositeExplicitAutograd, m) { + m.impl("get_workspace_size", TORCH_BOX(&get_workspace_size)); +} + +STABLE_TORCH_LIBRARY_IMPL(_flashkda_C, CUDA, m) { + m.impl("fwd", TORCH_BOX(&fwd)); +} REGISTER_EXTENSION(_flashkda_C)