Workaround custom_all_reduce issue (By Chengzhe Fan @zx3xyy)

[zjing14]

Note: the issue is identified and the fix is implemented by Chengzhe Fan @zx3xyy

Issue: Crash with custom all-reduce + CUDA graph enabled (ROCm / Aiter)

Root Cause Hypothesis: In Aiter's ROCm registered-input CUDA graph custom all-reduce path, a GPU can signal "ready" and peers begin reading the output buffer before the producer kernel's memory writes are globally visible across the PCIe/xGMI fabric.

Workaround: Insert a system-scope memory fence (e.g., __threadfence_system() or equivalent ROCm barrier) before the readiness signal, ensuring all prior stores from producer kernels within the CUDA graph are visible to peer GPUs before any consumer on another GPU reads the buffer.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 3290 --add-label <label>

[zjing14]

@HaiShaw @carlushuang