[PR #1189] [CLOSED] [NvlinkTransport] Fix HIP Shareable POSIX File Descriptor Handles

No labels

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

kvcache-ai/Mooncake#2062

bsakaguchi commented

2026-10-29 21:32:29 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/kvcache-ai/Mooncake/pull/1189
Author: @amd-arozanov
Created: 12/9/2025
Status: ❌ Closed

Base: main ← Head: posix_fd2

📝 Commits (5)

4628e81 [NvlinkTransport] Fix HIP shareable handles implementation
547b409 [NvlinkTransport] Fix hipMemGetAddressRange segmentation fault
13d58f3 Fix ROCm 6.x compatibility: remove requestedHandleTypes usage
be2dbcb [NvlinkTransport] Set proper device context for each transfer
18807f1 [NvlinkTransport] Cleanup

📊 Changes

3 files changed (+156 additions, -3 deletions)

View changed files

📝 mooncake-transfer-engine/include/gpu_vendor/hip.h (+5 -1)
📝 mooncake-transfer-engine/nvlink-allocator/nvlink_allocator.cpp (+7 -1)
📝 mooncake-transfer-engine/src/transport/nvlink_transport/nvlink_transport.cpp (+144 -1)

📄 Description

Description

(AI Assist)

This PR is focusing on fixes for proper POSIX file descriptor-based shareable handle implementation.

Changes:

Implement proper shareable handle support for HIP

Added open_fd() function that uses pidfd_open and pidfd_getfd syscalls to correctly open POSIX file descriptors from CUmemFabricHandle in remote processes

Fix `hipMemGetAddressRange` segmentation fault

Fixed segfault caused by passing NULL to hipMemGetAddressRange, which unlike CUDA's version doesn't support NULL arguments

Fix ROCm 6.x compatibility

Use requestedHandleType field instead of requestedHandleTypes to support ROCm 6.x
requestedHandleTypes support was added in ROCm 7.0

Add proper device context management

Extracted device ID from request.source to set correct device context for each transfer operation
Ensures memory operations execute on the appropriate GPU device

Type of Change

Types
- Bug fix
- New feature
  - Transfer Engine
  - Mooncake Store
  - Mooncake EP
  - Integration
  - P2P Store
  - Python Wheel
- Breaking change
- CI/CD
- Documentation update
- Other

How Has This Been Tested?

Ran transfer_engine_bench and transfer_engine_validator on nodes with AMG GPU

Checklist

I have performed a self-review of my own code.
I have updated the documentation.
I have added tests to prove my changes are effective.

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/kvcache-ai/Mooncake/pull/1189 **Author:** [@amd-arozanov](https://github.com/amd-arozanov) **Created:** 12/9/2025 **Status:** ❌ Closed **Base:** `main` ← **Head:** `posix_fd2` --- ### 📝 Commits (5) - [`4628e81`](https://github.com/kvcache-ai/Mooncake/commit/4628e81cd132bc5c5d698c9e2ec7b0bd64f1c0db) [NvlinkTransport] Fix HIP shareable handles implementation - [`547b409`](https://github.com/kvcache-ai/Mooncake/commit/547b40963dab8df16dfbf83b30b38877ed2abc48) [NvlinkTransport] Fix hipMemGetAddressRange segmentation fault - [`13d58f3`](https://github.com/kvcache-ai/Mooncake/commit/13d58f38c474c8123bd63dc92101a06a412e73a5) Fix ROCm 6.x compatibility: remove requestedHandleTypes usage - [`be2dbcb`](https://github.com/kvcache-ai/Mooncake/commit/be2dbcb7e45064033df448600c06117762ba0422) [NvlinkTransport] Set proper device context for each transfer - [`18807f1`](https://github.com/kvcache-ai/Mooncake/commit/18807f1f5855ed56eea31cfd4867d2d758971aa2) [NvlinkTransport] Cleanup ### 📊 Changes **3 files changed** (+156 additions, -3 deletions) <details> <summary>View changed files</summary> 📝 `mooncake-transfer-engine/include/gpu_vendor/hip.h` (+5 -1) 📝 `mooncake-transfer-engine/nvlink-allocator/nvlink_allocator.cpp` (+7 -1) 📝 `mooncake-transfer-engine/src/transport/nvlink_transport/nvlink_transport.cpp` (+144 -1) </details> ### 📄 Description ## Description (AI Assist) This PR is focusing on fixes for proper POSIX file descriptor-based shareable handle implementation. ### Changes: #### Implement proper shareable handle support for HIP - Added `open_fd()` function that uses `pidfd_open` and `pidfd_getfd` syscalls to correctly open POSIX file descriptors from `CUmemFabricHandle` in remote processes #### Fix `hipMemGetAddressRange` segmentation fault - Fixed segfault caused by passing NULL to `hipMemGetAddressRange`, which unlike CUDA's version doesn't support NULL arguments #### Fix ROCm 6.x compatibility - Use `requestedHandleType` field instead of `requestedHandleTypes` to support ROCm 6.x - `requestedHandleTypes` support was added in ROCm 7.0 #### Add proper device context management - Extracted device ID from `request.source` to set correct device context for each transfer operation - Ensures memory operations execute on the appropriate GPU device ## Type of Change * Types - [x] Bug fix - [ ] New feature - [x] Transfer Engine - [ ] Mooncake Store - [ ] Mooncake EP - [ ] Integration - [ ] P2P Store - [ ] Python Wheel - [ ] Breaking change - [ ] CI/CD - [ ] Documentation update - [ ] Other ## How Has This Been Tested? Ran transfer_engine_bench and transfer_engine_validator on nodes with AMG GPU ## Checklist - [x] I have performed a self-review of my own code. - [x] I have updated the documentation. - [x] I have added tests to prove my changes are effective. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

closed this issue
added the
pull-request
label

Rows
Columns

[PR #1189] [CLOSED] [NvlinkTransport] Fix HIP Shareable POSIX File Descriptor Handles #2062