减少lds使用, 提高并行度

csrc/sm90/decode/sparse_fp8/splitkv_mla.cuh

@@ -725,7 +725,7 @@ void KernelTemplate<MODEL_TYPE, NUM_HEADS>::run(const SparseAttnDecodeParams &pa
         KU_ASSERT(params.stride_kv_row == 656);  // number of bytes per token (512 fp8 + 4 float32 + 64 bfloat16)
     }
     auto mla_kernel = &flash_fwd_splitkv_mla_fp8_sparse_kernel<KernelTemplate<MODEL_TYPE, NUM_HEADS>>;
-    constexpr size_t smem_size = sizeof(SharedMemoryPlan);
+    constexpr size_t smem_size = 32768; // lds复用
     // zhj debug
     // printf("NUM_M_BLOCKS = %d smem_size = %d \n",NUM_M_BLOCKS, smem_size);
     mla_kernel<<<dim3(NUM_M_BLOCKS, params.s_q, params.num_sm_parts), NUM_THREADS, smem_size, params.stream>>>(params);