[Issue]: MORI-EP bug on MI300X+CX7

[TianDi101]

Problem Description

Stress test failed on MI300X+CX7 due to hang issue.
For benchmark, we met this error
/home/ditian12/mori/src/ops/dispatch_combine/internode_v1.cpp:375: void mori::moe::v1::DispatchInterNodeRecv(EpDispatchCombineArgs<T> &) [T = hip_bfloat16]: Device-side assertion (lanePe < config.worldSize) && (lanePe >= 0)' failed.`

Operating System

Ubuntu

CPU

AMD

GPU

MI300X

ROCm Version

ROCm-7.0.0

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

[JiakunYan]

Hi @TianDi101 , I ran into what seems like a related issue. SGLang + MORI hangs when running on two MI300X + CX7 nodes (inter_node mode), but works fine on a single node (intra_node mode). Is this expected behavior currently? Is anyone working on a fix?

I opened an issue with more details: sgl-project/sglang#19991