[perf] Speed up align sum kernels (vllm-project#21079)

Signed-off-by: Himanshu Jaju <[email protected]>

Author: hj-mistral

benchmarks/kernels/benchmark_moe_align_block_size.py

@@ -33,15 +33,13 @@ def check_correctness(num_tokens, num_experts=256, block_size=256, topk=8):
     sorted_ids_triton = torch.empty(
         (max_num_tokens_padded,), dtype=torch.int32, device="cuda"
     )
-    sorted_ids_triton.fill_(topk_ids.numel())  # fill with sentinel value
-    expert_ids_triton = torch.zeros(
+    expert_ids_triton = torch.empty(
         (max_num_tokens_padded // block_size,), dtype=torch.int32, device="cuda"
     )
     num_tokens_post_pad_triton = torch.empty((1,), dtype=torch.int32, device="cuda")
 
     sorted_ids_vllm = torch.empty_like(sorted_ids_triton)
-    sorted_ids_vllm.fill_(topk_ids.numel())
-    expert_ids_vllm = torch.zeros_like(expert_ids_triton)
+    expert_ids_vllm = torch.empty_like(expert_ids_triton)
     num_tokens_post_pad_vllm = torch.empty_like(num_tokens_post_pad_triton)
 
     # 2. run implementations
@@ -102,7 +100,6 @@ def benchmark(num_tokens, num_experts, topk, provider):
 
     max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
     sorted_ids = torch.empty((max_num_tokens_padded,), dtype=torch.int32, device="cuda")
-    sorted_ids.fill_(topk_ids.numel())
     max_num_m_blocks = max_num_tokens_padded // block_size
     expert_ids = torch.empty((max_num_m_blocks,), dtype=torch.int32, device="cuda")
     num_tokens_post_pad = torch.empty((1,), dtype=torch.int32, device="cuda")

csrc/moe/moe_align_sum_kernels.cu

@@ -1,6 +1,7 @@
 #include <torch/all.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
+#include <cub/cub.cuh>
 
 #include <ATen/ATen.h>
 #include <ATen/cuda/Atomic.cuh>
@@ -19,9 +20,14 @@ __global__ void moe_align_block_size_kernel(
     int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ expert_ids,
     int32_t* __restrict__ total_tokens_post_pad, int32_t num_experts,
     int32_t padded_num_experts, int32_t experts_per_warp, int32_t block_size,
-    size_t numel, int32_t* __restrict__ cumsum) {
+    size_t numel, int32_t* __restrict__ cumsum, int32_t max_num_tokens_padded) {
   extern __shared__ int32_t shared_counts[];
 
+  // Initialize sorted_token_ids with numel
+  for (size_t it = threadIdx.x; it < max_num_tokens_padded; it += blockDim.x) {
+    sorted_token_ids[it] = numel;
+  }
+
   const int warp_id = threadIdx.x / WARP_SIZE;
   const int my_expert_start = warp_id * experts_per_warp;
 
@@ -45,18 +51,27 @@ __global__ void moe_align_block_size_kernel(
 
   __syncthreads();
 
-  if (threadIdx.x == 0) {
-    cumsum[0] = 0;
-    for (int i = 1; i <= num_experts; ++i) {
-      int expert_count = 0;
-      int warp_idx = (i - 1) / experts_per_warp;
-      int expert_offset = (i - 1) % experts_per_warp;
-      expert_count = shared_counts[warp_idx * experts_per_warp + expert_offset];
+  // Compute prefix sum over token counts per expert
+  using BlockScan = cub::BlockScan<int32_t, 1024>;
+  __shared__ typename BlockScan::TempStorage temp_storage;
 
-      cumsum[i] =
-          cumsum[i - 1] + CEILDIV(expert_count, block_size) * block_size;
-    }
-    *total_tokens_post_pad = cumsum[num_experts];
+  int expert_count = 0;
+  int expert_id = threadIdx.x;
+  if (expert_id < num_experts) {
+    int warp_idx = expert_id / experts_per_warp;
+    int expert_offset = expert_id % experts_per_warp;
+    expert_count = shared_counts[warp_idx * experts_per_warp + expert_offset];
+    expert_count = CEILDIV(expert_count, block_size) * block_size;
+  }
+
+  int cumsum_val;
+  BlockScan(temp_storage).ExclusiveSum(expert_count, cumsum_val);
+  if (expert_id <= num_experts) {
+    cumsum[expert_id] = cumsum_val;
+  }
+
+  if (expert_id == num_experts) {
+    *total_tokens_post_pad = cumsum_val;
   }
 
   __syncthreads();
@@ -67,6 +82,13 @@ __global__ void moe_align_block_size_kernel(
       expert_ids[i / block_size] = threadIdx.x;
     }
   }
+
+  // Fill remaining expert_ids with 0
+  const size_t fill_start_idx = cumsum[num_experts] / block_size + threadIdx.x;
+  const size_t expert_ids_size = CEILDIV(max_num_tokens_padded, block_size);
+  for (size_t i = fill_start_idx; i < expert_ids_size; i += blockDim.x) {
+    expert_ids[i] = 0;
+  }
 }
 
 template <typename scalar_t>
@@ -105,7 +127,12 @@ __global__ void moe_align_block_size_small_batch_expert_kernel(
     const scalar_t* __restrict__ topk_ids,
     int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ expert_ids,
     int32_t* __restrict__ total_tokens_post_pad, int32_t num_experts,
-    int32_t block_size, size_t numel) {
+    int32_t block_size, size_t numel, int32_t max_num_tokens_padded) {
+  // Initialize sorted_token_ids with numel
+  for (size_t it = threadIdx.x; it < max_num_tokens_padded; it += blockDim.x) {
+    sorted_token_ids[it] = numel;
+  }
+
   const size_t tid = threadIdx.x;
   const size_t stride = blockDim.x;
 
@@ -153,6 +180,13 @@ __global__ void moe_align_block_size_small_batch_expert_kernel(
     }
   }
 
+  // Fill remaining expert_ids with 0
+  const size_t fill_start_idx = cumsum[num_experts] / block_size + threadIdx.x;
+  const size_t expert_ids_size = CEILDIV(max_num_tokens_padded, block_size);
+  for (size_t i = fill_start_idx; i < expert_ids_size; i += blockDim.x) {
+    expert_ids[i] = 0;
+  }
+
   for (size_t i = tid; i < numel; i += stride) {
     int32_t expert_id = topk_ids[i];
     int32_t rank_post_pad =
@@ -179,13 +213,17 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
   int threads = 1024;
   threads = ((threads + WARP_SIZE - 1) / WARP_SIZE) * WARP_SIZE;
 
+  // BlockScan uses 1024 threads and assigns one thread per expert.
+  TORCH_CHECK(padded_num_experts < 1024,
+              "padded_num_experts must be less than 1024");
+
   VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(
       topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
         // calc needed amount of shared mem for `cumsum` tensors
         auto options_int =
             torch::TensorOptions().dtype(torch::kInt).device(topk_ids.device());
         torch::Tensor cumsum_buffer =
-            torch::zeros({num_experts + 1}, options_int);
+            torch::empty({num_experts + 1}, options_int);
         bool small_batch_expert_mode =
             (topk_ids.numel() < 1024) && (num_experts <= 64);
 
@@ -203,7 +241,7 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
               sorted_token_ids.data_ptr<int32_t>(),
               experts_ids.data_ptr<int32_t>(),
               num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
-              topk_ids.numel());
+              topk_ids.numel(), sorted_token_ids.size(0));
         } else {
           auto align_kernel = vllm::moe::moe_align_block_size_kernel<scalar_t>;
 
@@ -217,7 +255,8 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
               experts_ids.data_ptr<int32_t>(),
               num_tokens_post_pad.data_ptr<int32_t>(), num_experts,
               padded_num_experts, experts_per_warp, block_size,
-              topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>());
+              topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>(),
+              sorted_token_ids.size(0));
 
           const int block_threads = std::min(256, (int)threads);
           const int num_blocks =

vllm/model_executor/layers/fused_moe/moe_align_block_size.py

@@ -111,6 +111,8 @@ def moe_align_block_size_triton(
                          dtype=torch.int32,
                          device=topk_ids.device)
     tokens_per_thread = cdiv(numel, num_experts)
+    sorted_token_ids.fill_(numel)
+    expert_ids.zero_()
 
     moe_align_block_size_stage1[grid](
         topk_ids,
@@ -205,11 +207,8 @@ def moe_align_block_size(
     sorted_ids = torch.empty((max_num_tokens_padded, ),
                              dtype=torch.int32,
                              device=topk_ids.device)
-    sorted_ids.fill_(topk_ids.numel())
     max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size)
-    # Expert ids must be zeroed out to prevent index out of bounds error while
-    # mapping global expert ids to local expert ids in expert parallelism.
-    expert_ids = torch.zeros((max_num_m_blocks, ),
+    expert_ids = torch.empty((max_num_m_blocks, ),
                              dtype=torch.int32,
                              device=topk_ids.device)
     num_tokens_post_pad = torch.empty((1),