[BUG] aclnnMoeDistributeDispatchV2 tiling failed on A2

[alanhe151220037]

When I enable deepep in sglang, error shows

[ERROR][MoeDistributeDispatchV2] Attr commAlg is invalid, current only support fullmesh_v1 and fullmesh_v2, but got commAlg = fullmesh.
[ERROR][MoeDistributeDispatchV2] Get special expert, commAlg attr and set tiling data failed.
[ERROR][MoeDistributeDispatchV2] Get attr and set tiling data failed.

The following code is self-contradictory

sgl-kernel-npu/csrc/deepep/deep_ep.cpp

Lines 796 to 800 in f057d77

	if (soc_version == op::SocVersion::ASCEND910B) {
	comm_alg = "fullmesh";
	} else {
	comm_alg = "fullmesh_v1";
	}

sgl-kernel-npu/csrc/deepep/ops/op_host/moe_distribute_dispatch_v2_tiling.cpp

Lines 604 to 610 in f057d77

	OP_TILING_CHECK(
	(strlen(commAlgPtr) != 0) && (strcmp(commAlgPtr, "fullmesh_v1") != 0) &&
	(strcmp(commAlgPtr, "fullmesh_v2") != 0),
	OP_LOGE(nodeName,
	"Attr commAlg is invalid, current only support fullmesh_v1 and fullmesh_v2, but got commAlg = %s.",
	commAlgPtr),
	return ge::GRAPH_FAILED);

[iforgetmyname]

@Yael-X would you please take a look at this?

[Yael-X]

@alanhe151220037 Please provide the version of cann you are using so that we can try to see if we can reproduce this issue.

[alanhe151220037]

@Yael-X Sorry, I found that I didn't use deepep2 to build for A2. But now I have another problem, I can't pass the UT run_test_internode.sh on 2 machines with 16 cards. It will stuck. (I do pass test_low_latency.py after I adapt the case on 2 machines with 16 cards whether I use the environment var HCCL_INTRA_PCIE_ENABLE=1 HCCL_INTRA_ROCE_ENABLE=0 or not.)

[Yael-X]

You can check out https://github.com/sgl-project/sgl-kernel-npu/blob/main/python/deep_ep/A2_DEEPEP_CN.md, which might be helpful.