diff --git a/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py b/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py index 5e950fdf60..916cebd74c 100644 --- a/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py +++ b/vllm/model_executor/layers/fused_moe/prepare_finalize/deepep_v2.py @@ -210,14 +210,23 @@ class DeepEPV2PrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeModular): else: expert_x, expert_x_scale = recv_x, None - expert_tokens_meta = mk.ExpertTokensMetadata.make_from_list( - recv_expert_num_tokens, - device=expert_x.device, - ) + if recv_expert_num_tokens: + expert_tokens_meta = mk.ExpertTokensMetadata.make_from_list( + recv_expert_num_tokens, + device=expert_x.device, + ) + else: + # Decode/cudagraph path (do_cpu_sync=False) skips the CPU sync and + # leaves recv_expert_num_tokens empty. A present-but-empty + # ExpertTokensMetadata violates the decode-mode contract above + # (expert_tokens_meta must be None) and crashes DeepEP combine + # during profile_run when CUDA graphs are enabled. + expert_tokens_meta = None if recv_topk_idx is None: # do_expand=True (prefill mode): build topk_ids from # per-expert token counts. + assert expert_tokens_meta is not None total_tokens = sum(recv_expert_num_tokens) if total_tokens > 0: recv_topk_idx = torch.repeat_interleave(