Getting started with AMD ROCm

Last updated: 05/17/2026.

Author: Mingjie Lu, Xiaohong Kou, Fuwei Yang

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: vLLM validated; SGLang support is ongoing.
Trainer backends: FSDP, FSDP2 and Megatron.
GPU targets:
- MI300X / MI325X (gfx942)
- MI355X (gfx950)

Software Baseline

Use the following prebuilt image for tutorial and validation:

amdagi/verl-dev:rocm7.0.2_56_te2.10_vllm0.20_py312

The Docker build recipe remains unchanged:

docker/rocm/Dockerfile.rocm

Host Prerequisites

Before launching the container, ensure:

AMD ROCm 7.0.2 host driver stack is installed and healthy.
Docker has access to /dev/kfd and /dev/dri.
Dataset and model storage paths are ready.

Launch Container

NAME=verl_release
DOCKER=amdagi/verl-dev:rocm7.0.2_56_te2.10_vllm0.20_py312

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)

# 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

Current support status
Category	Status	Notes
Runtime mode	Fully supported	Fully Async and Colocate are production-ready
Inference engine	vLLM validated	SGLang integration is ongoing
Trainer backend	Fully supported	FSDP, Megatron
Hardware	Fully supported	MI300X / MI325X (gfx942), MI355X (gfx950)

Example Workflow

1) Colocate mode + FSDP (GRPO, Qwen3-8B)

For Qwen3-8B FSDP training, enable both parameter and optimizer offload to avoid OOM.

# Configure these in your launch script or Hydra overrides:
# actor_rollout_ref.actor.fsdp_config.param_offload=True
# actor_rollout_ref.actor.fsdp_config.optimizer_offload=True
bash examples/grpo_trainer/run_qwen3_8b_fsdp.sh

2) Colocate mode + Megatron (GRPO, Qwen3.5-35B)

bash examples/grpo_trainer/run_qwen3_5-35b-megatron.sh

3) Fully Async mode

RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES and RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES are no longer required in this release.

# For qwen2.5-math-7b, update max_position_embeddings to 32768 in config.json after model download.
bash verl/experimental/fully_async_policy/shell/dapo_7b_math_fsdp2_4_4.sh