[GH-ISSUE #558] [Question] Why not check the CQE when ibgda_poll_cq #343

New issue

Open

opened 2026-10-26 09:58:52 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 09:58:52 +00:00

Owner

Originally created by @gangxie112 on GitHub (Jan 7, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/558

Hello,

We noticed that there is no CQE check in the DeepEP implemented ibgda_poll_cq, while nvshmem has the check as the following. Could some error be ignored, introducing some data corruption or hung?

if (unlikely(opcode == MLX5_CQE_REQ_ERR)) {
    ibgda_mlx5_err_cqe_t *cqe_err = (ibgda_mlx5_err_cqe_t *)cqe64;
    *error = cqe_err->syndrome;
    status = -1;
}

Originally created by @gangxie112 on GitHub (Jan 7, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/558 Hello, We noticed that there is no CQE check in the DeepEP implemented ibgda_poll_cq, while nvshmem has the check as the following. Could some error be ignored, introducing some data corruption or hung? if (unlikely(opcode == MLX5_CQE_REQ_ERR)) { ibgda_mlx5_err_cqe_t *cqe_err = (ibgda_mlx5_err_cqe_t *)cqe64; *error = cqe_err->syndrome; status = -1; }

bsakaguchi commented

2026-10-26 09:58:53 +00:00

Author

Owner

@gangxie112 commented on GitHub (Jan 21, 2026):

I think the reason is that:

any cqe error, according to ib spec, hca will transmit the state of the qp to error state
the dispatch/combine will not go further and get hung, then finally timed out.
There should be no data corruption issue. The dispatched/combined data will be discard by user, no corruption data will be consumed.

@gangxie112 commented on GitHub (Jan 21, 2026): I think the reason is that: 1. any cqe error, according to ib spec, hca will transmit the state of the qp to error state 2. the dispatch/combine will not go further and get hung, then finally timed out. 3. There should be no data corruption issue. The dispatched/combined data will be discard by user, no corruption data will be consumed.

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

deepseek-ai/DeepEP#343

No description provided.

Rows
Columns