Skip to content

[Model][Perf] Fuse Beam's TP all-reduce with its sandwich norms and MoE finalize #62366

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
WoosukKwon merged 6 commits into from
Oct 22, 2026
Merged

[Model][Perf] Fuse Beam's TP all-reduce with its sandwich norms and MoE finalize #62366

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
4fb4e75
[Model][Perf] Fuse Beam's TP all-reduce with its sandwich norms and M…
WoosukKwon Oct 18, 2026
5d68355
[Model][Perf] Build Beam's fused all-reduce ahead of time under csrc/…
WoosukKwon Oct 20, 2026
3ec2cf2
[Model][Perf] Note FlashInfer v0.7.2 as the vendored MNNVL source
WoosukKwon Oct 21, 2026
9fdc9ed
[Model][Perf] Keep Beam's deferred MoE finalize TP-only
WoosukKwon Oct 21, 2026
df583a4
[Build] Gate Beam's csrc kernels under one CMake switch
WoosukKwon Oct 21, 2026
7319588
Merge branch 'main' into woosuk/beam-fused-allreduce
WoosukKwon Oct 21, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 24 additions & 0 deletions CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1150,6 +1150,26 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
"Building fused GDN decode for archs: ${FUSED_GDN_DECODE_ARCHS}")
endif()

# Beam's kernels use NVLink multicast and thread-block clusters. Their norms
# must not contract into FMAs to keep the FP32 boundary exact.
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
cuda_archs_loose_intersection(BEAM_KERNEL_ARCHS
"9.0a;10.0f;10.3f" "${CUDA_ARCHS}")
else()
cuda_archs_loose_intersection(BEAM_KERNEL_ARCHS
"9.0a;10.0a;10.3a" "${CUDA_ARCHS}")
endif()
if(BEAM_KERNEL_ARCHS)
set(BEAM_KERNEL_SRCS "csrc/beam/allreduce_norm.cu")
set_gencode_flags_for_srcs(
SRCS "${BEAM_KERNEL_SRCS}"
CUDA_ARCHS "${BEAM_KERNEL_ARCHS}")
set_property(SOURCE ${BEAM_KERNEL_SRCS} APPEND PROPERTY
COMPILE_OPTIONS "$<$<COMPILE_LANGUAGE:CUDA>:--fmad=false>")
list(APPEND VLLM_STABLE_EXT_SRC "${BEAM_KERNEL_SRCS}")
message(STATUS "Building Beam kernels for archs: ${BEAM_KERNEL_ARCHS}")
endif()

if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
cuda_archs_loose_intersection(KIMI_K3_ATTN_RES_ARCHS
"10.0f" "${CUDA_ARCHS}")
Expand Down Expand Up @@ -1251,6 +1271,10 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
target_compile_definitions(_C_stable_libtorch PRIVATE
VLLM_ENABLE_KIMI_K3_ATTN_RES=1)
endif()
if(BEAM_KERNEL_ARCHS)
target_compile_definitions(_C_stable_libtorch PRIVATE
VLLM_ENABLE_BEAM_KERNELS=1)
endif()
# Needed by CUTLASS kernels
target_compile_definitions(_C_stable_libtorch PRIVATE
CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
Expand Down
Loading
Loading