verl_0720_main/verl/docker/rocm/README.md

ROCm (AMD GPU) Dockerfile of verl

This directory provides the Docker recipe for running verl on AMD GPUs with the ROCm software stack. The NVIDIA images described in ../README.md do not work on AMD hardware, so use Dockerfile.rocm instead.

For an end-to-end walkthrough (build, run, and example PPO/GRPO commands), see the tutorial: docs/amd_tutorial/amd_build_dockerfile_page.rst.

The other Dockerfile.rocm* / Apptainerfile.rocm files in this directory are kept only as historical references for older verl releases (ROCm 6.x, pinned verl 0.3.x / 0.4.x). New work should target Dockerfile.rocm.

Supported Hardware

The image targets the following GPU architectures (GPU_ARCH):

• gfx942 — MI300 series (MI300X / MI300A / MI325X)
• gfx950 — MI350 series (MI350X / MI355X)

Other architectures (e.g. gfx90a for MI200/MI250) can be built by overriding GPU_ARCH, but are not validated here.

Key Versions

Component	Version
ROCm	7.0.2
Python	3.12
PyTorch	2.9.1 (ROCm 7.0.2 wheel)
Triton	3.5.1
vLLM	source @ 1ff9d3353
Flash Attention	ROCm fork (CK backend) @ 83f9e450
TransformerEngine	ROCm fork @ 386bd316
aiter	ROCm @ 45c428e54
Megatron-core	0.16.0
Megatron-Bridge	0.5.0

What the Image Contains

Starting from a clean ubuntu:22.04 base, Dockerfile.rocm installs:

Prebuilt (downloaded), not compiled:

• ROCm 7.0.2 runtime + dev packages (via the repo.radeon.com apt repo)
• torch, apex, torchaudio, torchvision, triton — prebuilt ROCm wheels from repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/

Built from source (pinned commits):

• Flash Attention (ROCm fork, CK backend)
• TransformerEngine (ROCm fork)
• vLLM
• aiter

Also installed: cupy-rocm, mbridge, megatron-core, megatron-bridge, transformers, and the verl package itself.

Because Flash Attention / TransformerEngine / vLLM / aiter are compiled from source for the selected GPU architectures, the first build is slow (often 1-2+ hours). The image enables ccache (cached via a BuildKit cache mount) so that subsequent rebuilds are faster.

Building Locally

The Dockerfile uses BuildKit cache mounts (RUN --mount=...), so BuildKit is required (with the buildx plugin). If you see the --mount option requires BuildKit, install buildx (sudo apt-get install -y docker-buildx) and prefix the build with DOCKER_BUILDKIT=1.

DOCKER_BUILDKIT=1 docker build \
    -f docker/rocm/Dockerfile.rocm \
    -t verl-rocm:local .

Useful build arguments

Build arg	Default	Purpose
GPU_ARCH	gfx942;gfx950	GPU architectures to compile kernels for. Set to a single arch (e.g. gfx942) to roughly halve Flash Attention build time.
ROCM_VERSION / AMDGPU_VERSION	7.0.2	ROCm / amdgpu apt repo version. Note: the prebuilt torch/triton/etc. wheel URLs in the Dockerfile are pinned to ROCm 7.0.2; changing this also requires updating those URLs.
PYTHON_VERSION	3.12	Python version. Note: the prebuilt wheel URLs are pinned to the cp312 ABI; changing this also requires updating those URLs.
MAX_JOBS	$(nproc)	Parallel compile jobs. Lower it (e.g. 64) if the vLLM build runs out of memory.
FA_TAG / TE_TAG / VLLM_TAG / AITER_TAG	pinned	Source commits for the from-source components.

Example — build only for MI300 with a memory-safe job count:

DOCKER_BUILDKIT=1 docker build \
    -f docker/rocm/Dockerfile.rocm \
    --build-arg GPU_ARCH=gfx942 \
    --build-arg MAX_JOBS=64 \
    -t verl-rocm:mi300 .

Release History

• 2026/06/03: ROCm 7.0.2 stack — torch==2.9.1, triton==3.5.1, vLLM @1ff9d3353, Flash Attention (CK) @83f9e450, TransformerEngine @386bd316, aiter @45c428e54, megatron-core==0.16.0; targets gfx942 / gfx950.