[ROCm] [CI] Add ROCm Dockerfile and AMD CI (vllm v0.12.0)

[tjtanaa]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

Setting up CI and AMD Docker images

AMD CI

The vLLM Omni AMD CI is setup following the vLLM AMD CI. Given that the whole setup (agents and agent environment) is the same, I have adapted the bootstrap files from vLLM AMD CI to make it work with vLLM Omni AMD CI.

I suggest that we keep the bootstrap code confined in this repo first for easier maintainability. Let me know if anyone thinks otherwise.

The difference between the vLLM Omni AMD CI and vLLM AMD CI is that we try to make everything upstream compatible. The AMD CI docker image is store at the same aws repo as the other hardware.

The tests that are currently enabled:

• "Diffusion Model Test"
• "Diffusion Cache Backend Test"
• "Omni Model Test Qwen2-5-Omni"

The remaining test is under investigation:

• "Omni Model Test Qwen3-Omni"

AMD Docker images

Build docker image. The docker image is also used for CI.

docker build -f docker/Dockerfile.rocm -t vllm-omni-rocm .

How to launch the docker image

#!/bin/bash
docker run -it \
   --privileged \
   --network=host \
   --group-add=video \
   --ipc=host \
   --cap-add=SYS_PTRACE \
   --security-opt seccomp=unconfined \
   --device /dev/kfd \
   --device /dev/dri \
   --name vllm-omni-rocm \
   vllm-omni-rocm \
   bash

Update installation step to include docker build instruction and docker launch instruction.

Test Plan

Built locally and is able to run the examples without failure.

Passed the vllm-omni AMD CI.
Latest Run: https://buildkite.com/vllm/vllm-omni-amd-ci/builds/43/steps/canvas You can also find it in the CI check at the bottom of this page.

Test Result

Ran the Qwen-image examples

cd examples/offline_inference/qwen_image/

python text_to_image.py \
  --prompt "a cup of coffee on the table" \
  --seed 42 \
  --cfg_scale 4.0 \
  --num_images_per_prompt 1 \
  --num_inference_steps 50 \
  --height 1024 \
  --width 1024 \
  --output outputs/coffee.png

Extra details for future PR

create_new_process_for_each_test decorator does not seem to be working. When it is in spawn mode, the content of the test decorated by create_new_process_for_each_test is not going to run, and the test returned as PASSED immediately.
The tests on CUDA and NPU are able to run correctly because on these two platforms, the create_new_process_for_each_test will be evaluated as using fork and vLLM internally is running as spawn.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)

[hongxiayang]

Please simply this.

Suggested change

	ARG ROCM_BASE_IMAGE=rocm/vllm-dev
	ARG ROCM_BASE_TAG=nightly_main_20251005
	ARG BASE_IMAGE=rocm/vllm-dev:nightly_main_20251005

[hongxiayang]

Suggested change

	FROM ${ROCM_BASE_IMAGE}:${ROCM_BASE_TAG}
	FROM ${BASE_IMAGE}

[hongxiayang]

Please use the same /app directory as original vllm image

Suggested change

	ARG APP_DIR=/workspace/vllm-omni
	ARG COMMON_WORKDIR=/app

[hongxiayang]

Suggested change

	WORKDIR ${APP_DIR}
	WORKDIR ${COMMON_WORKDIR}

[hongxiayang]

Suggested change

	RUN cd ../ && python3 -m pip uninstall -y vllm && \
	RUN python3 -m pip uninstall -y vllm && \

[hongxiayang]

we should not put this in the common Dockerfile.

Suggested change

	ENV VLLM_ROCM_USE_AITER=1
	ENV VLLM_ROCM_USE_AITER_MHA=1
	ENV VLLM_ROCM_USE_AITER_LINEAR=0
	ENV VLLM_ROCM_USE_AITER_RMSNORM=0

[hongxiayang]

These environment variables, if needed, can be set when invoking/running the unit tests.

[hongxiayang]

adjust to use {WORKDIR}

[tjtanaa]

(18 Dec 2025) Passed https://buildkite.com/vllm/vllm-omni-amd-ci/builds/37/steps/canvas

[tjtanaa]

This PR also resolves #387

[tjtanaa]

@hongxiayang your requested review has been addressed.

@hsliuustc0106 Could you help to review this PR? Thank you. The unit tests for the Teacache and DiTCache are also passing.

[ZJY0516]

Could you explain why we don't need this in https://github.com/vllm-project/vllm/blob/ff2168bca3a195b835c64a5c9012d7b6a9f34e61/vllm/model_executor/layers/rotary_embedding/common.py#L122

[tjtanaa]

You are right. We can simplify the conditions and not use is_rocm() follow the implementation on vLLM side. Initially, I added is_rocm so that other platform can handle the case themselves. I do not know what are the libraries the other platforms have. Like how we missed the path because authors are not aware of what library does other platform has.

[tjtanaa]

@ZJY0516 Sorry. I don't know if I understand your question correctly, this is a new abstraction https://github.com/vllm-project/vllm/blob/ff2168bca3a195b835c64a5c9012d7b6a9f34e61/vllm/model_executor/layers/rotary_embedding/common.py#L122 that is after vllm v0.12.0. If we want to adopt this code, I suggest we do this in another PR.