permute/unpermute kernel for moe optimization (vllm-project#14568)

Signed-off-by: Caleb_Du <Caleb_Du@zju.edu.cn>
Signed-off-by: minpeter <kali2005611@gmail.com>

Author: CalebDu

CMakeLists.txt

@@ -15,7 +15,6 @@ project(vllm_extensions LANGUAGES CXX)
 
 # CUDA by default, can be overridden by using -DVLLM_TARGET_DEVICE=... (used by setup.py)
 set(VLLM_TARGET_DEVICE "cuda" CACHE STRING "Target device backend for vLLM")
-
 message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
 message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
 
@@ -682,6 +681,17 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 endif()
 
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  set(MOE_PERMUTE_SRC
+      "csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.cu"
+      "csrc/moe/moe_permute_unpermute_op.cu")
+
+  set_gencode_flags_for_srcs(
+    SRCS "${MARLIN_PERMUTE_SRC}"
+    CUDA_ARCHS "${MOE_PERMUTE_ARCHS}")
+
+  list(APPEND VLLM_MOE_EXT_SRC "${MOE_PERMUTE_SRC}")
+endif()
 message(STATUS "Enabling moe extension.")
 define_gpu_extension_target(
   _moe_C
@@ -690,6 +700,8 @@ define_gpu_extension_target(
   SOURCES ${VLLM_MOE_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
   USE_SABI 3
   WITH_SOABI)
 

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -90,7 +90,8 @@ def bench_run(results: list[benchmark.Measurement], model: str,
 
     score = torch.randn((m, num_experts), device="cuda", dtype=dtype)
 
-    topk_weights, topk_ids = fused_topk(a, score, topk, renormalize=False)
+    topk_weights, topk_ids, token_expert_indices = fused_topk(
+        a, score, topk, renormalize=False)
 
     def run_triton_moe(a: torch.Tensor, w1: torch.Tensor, w2: torch.Tensor,
                        topk_weights: torch.Tensor, topk_ids: torch.Tensor,

benchmarks/kernels/benchmark_moe.py

@@ -115,8 +115,8 @@ def run():
         from vllm.model_executor.layers.fused_moe import override_config
         with override_config(config):
             if use_deep_gemm:
-                topk_weights, topk_ids = fused_topk(x, input_gating, topk,
-                                                    False)
+                topk_weights, topk_ids, token_expert_indices = fused_topk(
+                    x, input_gating, topk, False)
                 return fused_experts(
                     x,
                     w1,