fix moe_align_block_size (#2615)

Author: HandH1998

sgl-kernel/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sgl-kernel"
-version = "0.0.2.post9"
+version = "0.0.2.post10"
 description = "Kernel Library for SGLang"
 readme = "README.md"
 requires-python = ">=3.8"

sgl-kernel/src/sgl-kernel/csrc/moe_align_kernel.cu

@@ -118,31 +118,19 @@ __global__ void moe_align_block_size_kernel(scalar_t* __restrict__ topk_ids, int
 }
 
 void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts, int64_t block_size,
-                          torch::Tensor sorted_token_ids, torch::Tensor experts_ids,
-                          torch::Tensor num_tokens_post_pad) {
+                          torch::Tensor sorted_token_ids, torch::Tensor experts_ids, torch::Tensor num_tokens_post_pad,
+                          torch::Tensor token_cnts_buffer, torch::Tensor cumsum_buffer) {
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   DISPATCH_INTEGRAL_TYPES(topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
     // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
     // tensors
     const int32_t num_thread = max((int32_t)num_experts, WARP_SIZE);
 
-    const int32_t mem_tokens_cnts = ((num_experts + 1) * num_experts) * sizeof(int32_t);
-    const int32_t mem_cumsum = (num_experts + 1) * sizeof(int32_t);
-
-    // allocate global memory
-    int32_t* tokens_cnts;
-    int32_t* cumsum;
-    cudaMalloc(&tokens_cnts, mem_tokens_cnts);
-    cudaMalloc(&cumsum, mem_cumsum);
-
-    // set dynamic shared mem
     auto kernel = moe_align_block_size_kernel<scalar_t>;
     kernel<<<1, num_thread, 0, stream>>>(topk_ids.data_ptr<scalar_t>(), 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(), tokens_cnts, cumsum);
-
-    cudaFree(tokens_cnts);
-    cudaFree(cumsum);
+                                         num_experts, block_size, topk_ids.numel(),
+                                         token_cnts_buffer.data_ptr<int32_t>(), cumsum_buffer.data_ptr<int32_t>());
   });
 }
 

sgl-kernel/src/sgl-kernel/ops/__init__.py

@@ -8,6 +8,8 @@ def moe_align_block_size(
     sorted_token_ids,
     experts_ids,
     num_tokens_post_pad,
+    token_cnts_buffer,
+    cumsum_buffer,
 ):
     _moe_align_block_size(
         topk_ids,
@@ -16,4 +18,6 @@ def moe_align_block_size(
         sorted_token_ids,
         experts_ids,
         num_tokens_post_pad,
+        token_cnts_buffer,
+        cumsum_buffer,
     )

sgl-kernel/tests/test_moe_align.py

@@ -18,8 +18,22 @@ def test_moe_align_block_size():
     )
     num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
 
+    token_cnts_buffer = torch.empty(
+        (num_experts + 1) * num_experts, dtype=torch.int32, device=topk_ids.device
+    )
+    cumsum_buffer = torch.empty(
+        num_experts + 1, dtype=torch.int32, device=topk_ids.device
+    )
+
     moe_align_block_size(
-        topk_ids, num_experts, block_size, sorted_ids, expert_ids, num_tokens_post_pad
+        topk_ids,
+        num_experts,
+        block_size,
+        sorted_ids,
+        expert_ids,
+        num_tokens_post_pad,
+        token_cnts_buffer,
+        cumsum_buffer,
     )