Getting started with AMD ROCm

Last updated: 07/24/2026.

Author: Mingjie Lu <https://github.com/mingjielu>, Xiaohong Kou <https://github.com/xiaohong42>, Fuwei Yang <https://github.com/amd-fuweiy>, Zhaodong Bing <https://github.com/aaab8b>

Overview

This document is a quick-start tutorial for running VeRL on AMD ROCm. It provides a production-style bring-up flow for container startup, environment verification, and training examples.

Current software and hardware scope:

• Runtime modes: fully supports Fully Async and Colocate.

• Inference engine: fully supports vLLM and SGLang.

• Trainer backends: FSDP, FSDP2 and Megatron.

• GPU targets:

  • MI300X / MI325X (gfx942)
  • MI350X / MI355X (gfx950)

Software Baseline

Use the following prebuilt image for tutorial and validation:

• amdagi/verl-dev:rocm7.14_torch2.12_release_0724

Or build from source:

• docker/rocm/Dockerfile.rocm <https://github.com/verl-project/verl/blob/main/docker/rocm/Dockerfile.rocm>_

Host Prerequisites

Before launching the container, ensure:

1. AMD ROCm 7.14 host driver stack is installed and healthy.
2. Docker has access to /dev/kfd and /dev/dri.
3. Dataset and model storage paths are ready.

Launch Container

.. code-block:: bash

NAME=verl_release
DOCKER=amdagi/verl-dev:rocm7.14_torch2.12_release_0724

docker pull $DOCKER

docker run -it --name $NAME --device /dev/kfd --device /dev/dri \
  --privileged --network=host \
  --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined \
  --shm-size=2048g \
  --ulimit memlock=-1 --ulimit stack=67108864 \
  -w /workspace \
  $DOCKER \
  /bin/bash

Environment Check (Inside Container)

.. code-block:: bash

# ROCm and visible GPU targets
rocminfo | grep -E "gfx942|gfx950" || true

# PyTorch + ROCm sanity check
python - <<'PY'
import torch
print("torch:", torch.__version__)
print("rocm :", torch.version.hip)
print("cuda_available:", torch.cuda.is_available())
if torch.cuda.is_available():
    print("gpu_count:", torch.cuda.device_count())
    print("device_0:", torch.cuda.get_device_name(0))
PY

Feature Support Matrix

.. list-table:: Current support status :header-rows: 1

• • Category
  • Notes
• • Runtime mode
  • Fully Async, Colocate
• • Inference engine
  • vLLM, SGLang
• • Trainer backend
  • FSDP, FSDP2, Megatron
• • Hardware
  • MI300X / MI325X (gfx942), MI350X/MI355X (gfx950)

Example Workflow

.. list-table:: :header-rows: 1

• • Runtime mode
  • Inference engine
  • Trainer backend
  • Example scripts
• • Colocate
  • vLLM
  • FSDP
  • bash examples/grpo_trainer/run_qwen3_8b_fsdp.sh <../../examples/grpo_trainer/run_qwen3_8b_fsdp.sh>_
• • Colocate
  • vLLM
  • Megatron
  • bash examples/grpo_trainer/run_qwen3_5_35b_megatron.sh <../../examples/grpo_trainer/run_qwen3_5_35b_megatron.sh>_
• • Fully Async
  • vLLM
  • FSDP2
  • bash verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_4_4.sh <../../verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_4_4.sh>_
• • Fully Async
  • vLLM
  • Megatron
  • bash verl/experimental/fully_async_policy/shell/geo3k_qwen25vl_7b_megatron_4_4.sh <../../verl/experimental/fully_async_policy/shell/geo3k_qwen25vl_7b_megatron_4_4.sh>_
• • Colocate
  • SGLang
  • FSDP
  • bash examples/grpo_trainer/run_qwen3_8b_fsdp.sh <../../examples/grpo_trainer/run_qwen3_8b_fsdp.sh>_ (vllm->sglang)
• • Colocate
  • SGLang
  • Megatron
  • bash examples/grpo_trainer/run_qwen3_8b_megatron.sh <../../examples/grpo_trainer/run_qwen3_8b_megatron.sh>_ (vllm->sglang)
• • Fully Async
  • SGLang
  • FSDP2
  • bash verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_4_4.sh <../../verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_4_4.sh>_ (vllm->sglang)
• • Fully Async
  • SGLang
  • Megatron
  • bash verl/experimental/fully_async_policy/shell/geo3k_qwen25vl_7b_megatron_4_4.sh <../../verl/experimental/fully_async_policy/shell/geo3k_qwen25vl_7b_megatron_4_4.sh>_ (vllm->sglang)

Known Issues

1. For qwen2.5-math-7b, update max_position_embeddings to 32768 in config.json after model download.
2. PYTORCH_ALLOC_CONF=expandable_segments:True is set by default in Dockerfile.rocm to prevent out-of-memory (OOM). However, this setting may conflict with vllm_custom_all_reduce, so we set vllm.disable_custom_all_reduce=True in config.yaml by default which will be removed in the future once ROCm resolves the conflict.
3. attention_backend must be set to triton for SGLang.
4. Several ROCm-specific pull requests for vLLM and SGLang are still pending upstream integration, so we currently incorporate the required patches directly within the Dockerfile. In the future, we intend to transition to stable, officially released versions.