[1/N][port from deepseek085] add custom allreduce from AITER

[zejunchen-zejun]

Sync deepseek085 optimization to rocm/vllm llama branch.
Will upstream same code changes to public vllm.

The custom allreduce is controlled by VLLM_ROCM_USE_AITER_CUSTOM_ALL_REDUCE(default: True).
If AITER is imported, the custom allreduce will be default used.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[zejunchen-zejun]

Here is the accuracy verification. Use Qwen/Qwen3-0.6B and gsm8k dataset.
Use CUDA allreduce:

Use AITER allreduce(export VLLM_ROCM_USE_AITER_CUSTOM_ALL_REDUCE=1):

No accuracy degradation is seen.

Verify command:

export VLLM_ROCM_USE_AITER_CUSTOM_ALL_REDUCE=1

#!/bin/bash
rm -rf /root/.cache/vllm
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

VLLM_USE_V1=1 \
VLLM_ROCM_USE_AITER=1 \
SAFETENSORS_FAST_GPU=1 \
lm_eval --model vllm --model_args pretrained=Qwen/Qwen3-0.6B,tensor_parallel_size=8,max_model_len=10000,gpu_memory_utilization=0.2 --trust_remote_code --tasks gsm8k --batch_size auto 2>&1 | tee ./pr_gsm8k-Qwen_Qwen3-32B-aiter-v1-3.log

[tjtanaavllm]

LGTM @zejunchen-zejun Thank you for the feature.