[rollout, sglang] feat: support sglang ROCm backend via aiter defaults and ray init env

[xiaohong42]

What does this PR do?

Enable the SGLang rollout backend to run out-of-the-box on AMD ROCm GPUs.

On ROCm (torch.version.hip is not None), this PR:

1. Defaults the SGLang attention backend to aiter when the user has not explicitly set one (previously it fell through to flashinfer/fa3).
2. Defaults SGLANG_USE_AITER=1 via os.environ.setdefault, so the non-attention kernels (RMSNorm / RoPE / MoE / quant) also take the AITER path. This is decoupled from the attention backend selection and is set for ROCm regardless of which attention backend is chosen.
3. Injects RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0 into PPO_RAY_RUNTIME_ENV so Ray 2.x does not force-clear accelerator visibility for the num_gpus=0 SGLang server actor (otherwise the actor sees no GPU).

All three are scoped to ROCm and remain user-overridable: an explicit engine_kwargs.sglang.attention_backend or an exported SGLANG_USE_AITER still win, and CUDA / NPU behavior is unchanged.

Checklist Before Starting

• Search for similar PRs. Query links:
  • https://github.com/volcengine/verl/pulls?q=is%3Apr+is%3Aopen+SGLANG_USE_AITER (0 results)
  • https://github.com/volcengine/verl/pulls?q=is%3Apr+RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO (no related PR)
• Format the PR title as [{modules}] {type}: {description}
  • Title: [rollout, sglang] feat: support sglang ROCm backend via aiter defaults and ray init env
  • Not a breaking change: defaults only apply when unset, and existing explicit settings keep working.

Test

Validated on AMD ROCm (MI3xx). Both rollout integration paths train end-to-end:

• Colocate (main_ppo, GRPO + SGLang): ran launch_smoke_qwen3_8b.sh.
• Fully async (fully_async_policy, SGLang): ran dapo_7b_math_fsdp2_4_4.sh.

Both smoke runs complete the training steps successfully. The SGLang server log shows the expected ROCm detected: attention_backend=aiter, SGLANG_USE_AITER=1.

CI does not cover this change because it requires ROCm hardware (AITER kernels + HIP visibility); the existing CPU/CUDA CI cannot exercise the ROCm path.

API and Usage Example

No API changes. The new behavior is automatic on ROCm and overridable:

# Default on ROCm: attention_backend=aiter, SGLANG_USE_AITER=1 (auto)

# Override the attention backend (SGLANG_USE_AITER still defaults to 1):
... +actor_rollout_ref.rollout.engine_kwargs.sglang.attention_backend=triton

# Fall back to vLLM RMSNorm/RoPE kernels by disabling AITER explicitly:
export SGLANG_USE_AITER=0

Design & Code Changes

• verl/workers/rollout/sglang_rollout/async_sglang_server.py
  • ROCm -> default attention_backend="aiter" when unset.
  • Set SGLANG_USE_AITER=1 via setdefault, decoupled from the attention backend, plus an info log of the resolved values.
• verl/trainer/constants_ppo.py
  • On ROCm, add RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0 to PPO_RAY_RUNTIME_ENV (conditional dict, mirroring the existing _gb200_nccl_env pattern).

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks (pre-commit run --files <changed files> --show-diff-on-failure --color=always): all hooks pass.
• Add / Update the documentation. — No docs change; behavior is auto + overridable and described above.
• Add unit or end-to-end test(s) ... If not feasible, explain why: requires ROCm hardware; validated manually (see Test).
• Once your PR is ready for CI, send a message in the ci-request channel. — To be done when opening the upstream PR.
• Not related to the recipe submodule.