Skip to content

fix: restore rdmaRegister guard and guard heapRkeys_ memcpy in proxy path - #18

Merged
itej89 merged 1 commit into
feat/ep-rdma-sharingfrom
feat/ep-rdma-sharing-fix-symm-mem
Aug 27, 2026
Merged

fix: restore rdmaRegister guard and guard heapRkeys_ memcpy in proxy path#18
itej89 merged 1 commit into
feat/ep-rdma-sharingfrom
feat/ep-rdma-sharing-fix-symm-mem

Conversation

@itej89

@itej89 itej89 commented Aug 27, 2026

Copy link
Copy Markdown
Owner

Summary

Addresses PR ROCm#558 review comments #5 and #6.

  1. feat: CPU proxy RDMA for rail-isolated EP (v3) #6: Restore rdmaRegister check in proxy path — without it, SDMA-only buffers hit ibv_reg_mr which can fail on ionic's single-MR limit.
  2. EP over RDMA v6: TransportType::PROXY — native dispatch framework #5: Guard heapRkeys_ memcpy against empty vector — in ShmemMode::Isolation the heap is never registered, so heapRkeys_ is empty and .data() returns nullptr.

Test plan

  • EP benchmark on DO AINIC: Dispatch Pass, Combine Pass
  • Dispatch avg 31.85 GB/s, Combine avg 58.01 GB/s (consistent)

🤖 Generated with Claude Code

…path

1. Restore rdmaRegister check in proxy path (line 198) — without it,
   SDMA-only buffers that shouldn't be registered hit ibv_reg_mr,
   which can fail on ionic's single-MR limit.

2. Guard heapRkeys_ memcpy against empty vector — in ShmemMode::Isolation
   the heap is never registered so heapRkeys_ is empty, causing memcpy
   from nullptr.

Addresses PR ROCm#558 review comments #5 and #6.

Co-Authored-By: Claude <noreply@anthropic.com>
@itej89
itej89 merged commit b008bf2 into feat/ep-rdma-sharing Aug 27, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant