[ZJLi2013]

Purpose

Adds a ROCm recipe for DreamZero-DROID. recipes/ currently covers one robot policy, GR00T-N1.7, CUDA only.

Two things about the published ROCm images, both recorded in the recipe's Notes:

• v0.24.0 and v0.24.1 cannot serve DreamZero — their dreamzero_tp1_cfg2.yaml is missing engine_backend, added in [Bugfix] Select the AR-Diffusion engine in the DreamZero CFG-parallel deploy config #5439, but not in a published ROCM image yet. here still with v0.22.0
• v0.22.0 needs pip install openpi-client on the server, which the image does not ship. The failure looks like a client-side ConnectionClosedError, so the note points at the server log. Not needed after v0.22.0.

Test Plan

vLLM Version: 0.22.0+rocm722

vLLM-Omni Commit: base d77a35a3; the documented configuration is tag v0.22.0

2x AMD MI300X (gfx942), ROCm/HIP 7.2.53211, Ubuntu 22.04.5, Python 3.12, torch 2.10.0+git8514f05, image
vllm/vllm-omni-rocm:v0.22.0@sha256:be06edae7ba9d14a89c5907da2e3631b7998765bce285c937f7d9a96e6aeb2c5.

On a clean v0.22.0 checkout with the published image — no locally built image, no out-of-tree patches:

1. The e2e test:

python -m pytest -q \
    tests/e2e/online_serving/test_dreamzero_expansion.py::test_dreamzero_openpi_online

2. The recipe's vllm serve command, driven by the in-tree example client
   examples/online_serving/dreamzero/openpi_client.py.

Test Result

1 passed, 17 warnings in 184.01s (0:03:04)

Server metadata: {"action_space": "joint_position", "image_resolution": [180, 320],
                  "n_external_cameras": 2, "needs_session_id": true,
                  "needs_stereo_camera": false, "needs_wrist_camera": true}
Action 0: shape=(24, 8) dtype=float32 min=-0.120401 max=0.540639
Action 1: shape=(24, 8) dtype=float32 min=-0.003906 max=0.421654

Weights load at 42.79 GiB per rank.

Not in scope

• No latency numbers. Functional validation only.
• Single session. max_num_seqs: 1; stated as a limitation, not tested.
• ROCm section only, leaving the CUDA section to whoever validates it.
• Not validated against main. v0.22.0 is the newest ROCm image published.
  Re-pin when a newer one ships; that also retires both notes above.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[ZJLi2013]

mug_dwell_r1.mp4

[vllm-omni-review-bot]

Omni ReviewBot: no human activity for 25 days

@ZJLi2013 this pull request has had no human commit, comment or review since 2026-08-14. Per repository policy it may be closed if it stays inactive.

To keep it moving, any one of these is enough: push an update, reply to the open blocker, or post the current plan and timeline.

[vllm-omni-review-bot]

Omni ReviewBot: no human activity for 40 days

@ZJLi2013 this pull request has had no human commit, comment or review since 2026-08-14. Please consider closing this PR if the work is no longer planned. The author or a maintainer decides whether to change the PR state.

To keep it moving, any one of these is enough: push an update, reply to the open blocker, or post the current plan and timeline.

[hsliuustc0106]

@ZJLi2013 Are you planning to continue work on this PR? Please reply by 2026-10-19 10:43:56 UTC.

A yes or a brief plan is enough; code changes do not need to be finished within these three days. If you are waiting on a review or other maintainer action, please let us know.

If there is no reply, new commit, or maintainer takeover by then, we will recheck and close this PR as inactive. It can be reopened when work resumes.