diff --git a/vllm/models/deepseek_v4/nvidia/flashmla.py b/vllm/models/deepseek_v4/nvidia/flashmla.py index 29ec2d6f2a..f0b00163d1 100644 --- a/vllm/models/deepseek_v4/nvidia/flashmla.py +++ b/vllm/models/deepseek_v4/nvidia/flashmla.py @@ -96,8 +96,11 @@ class DeepseekV4FlashMLAAttention(DeepseekV4Attention): // self.compress_ratio ) M = N + self.window_size + self.max_num_batched_tokens - assert self.topk_indices_buffer is not None - top_k = 0 if swa_only else self.topk_indices_buffer.shape[-1] + if swa_only: + top_k = 0 + else: + assert self.topk_indices_buffer is not None + top_k = self.topk_indices_buffer.shape[-1] combined_topk = round_up(top_k + self.window_size, 128) current_workspace_manager().get_simultaneous( ((self.PREFILL_CHUNK_SIZE, M, q.shape[-1]), torch.bfloat16),