使用64位计算stride,避免越界

csrc/gfx93/decode/sparse_fp8/splitkv_mla.cuh

@@ -366,7 +366,7 @@ __device__ void KernelTemplate<MODEL_TYPE, NUM_HEADS>::compute_attn_1rowblock_sp
         }
         int block_index = token_index == -1 ? 0 : (int)((uint32_t)token_index/(uint32_t)page_block_size);   // Use uint32_t division and mod to improve performance        const int token_indexrel_idx_in_block = (token_index + page_block_size) % page_block_size;
         int rel_idx_in_block = (uint32_t)token_index % (uint32_t)page_block_size;   // NOTE When token_index is -1 (UINT_MAX), UINT_MAX%page_block_size < page_block_size, so there will be no illegal-memory-access error
-        const index_t offset_k = block_index * k_block_stride;
+        const index_t offset_k = (index_t)(block_index) * k_block_stride;
         uint8_t* gK_base;
         float scales[NUM_SCALES];
         if constexpr (MODEL_TYPE == ModelType::V32) {