[Feature]: Suggestions Regarding IBGDA Performance

Suggestion Description

Hello, I've briefly compared deepseek-ai/DeepEP and ROCm/DeepEP and have the following questions:

1. deepseek-ai/DeepEP implements its own low-latency IGBD process, avoiding polling CQ when issuing WQEs. The official response is: https://github.com/deepseek-ai/DeepEP/issues/180. ROCm/DeepEP, however, directly calls the API provided by rocshmem. Each time a WQE is issued, it checks for available space; if not, it polls CQ. This approach is essentially the same as the IGBD process implemented by nvshmem, potentially leading to higher latency in low-latency mode.
2. ROCm/DeepEP calls a warp interface similar to put_nbi_warp. In rocshmem, only one thread actually issues WQEs, while in deepseek-ai/DeepEP, all threads participate in the warp. Wouldn't this affect performance?

Operating System

No response

GPU

No response

ROCm Component

No response

Review the reported differences between ROCm/DeepEP and deepseek-ai/DeepEP, especially rocshmem's WQE and CQ polling behavior and warp participation. Compare the referenced DeepEP issue and determine whether a concrete performance change is warranted; done requires an agreed scope and evidence that the low-latency behavior improves.