[panpan0000]

When I was digging which image works best for Minimax-H3, I found this issue on Dockerfile...

This PR are created with AI assistant...

Summary

• docker/Dockerfile.cuda (the Dockerfile built by .buildkite/release/release-pipeline.yml for release and nightly CUDA images) still defaults to ARG BASE_IMAGE=vllm/vllm-openai:v0.26.0, while the codebase was rebased onto vLLM 0.27.0 in [Rebase] Rebase to vllm 0.27.0 #5976.
• Result: the published vllm/vllm-omni:v0.27.0rc1, latest, and daily nightly CUDA images install current vLLM-Omni — which uses vLLM 0.27-only APIs — over a vLLM 0.26.0 runtime, a combination CI never validates (docker/Dockerfile.ci uses VLLM_BASE_TAG=v0.27.0).
• This one-line change aligns the release image base with the CI image base, matching the rebase rule noted in Dockerfile.ci ("Keep VLLM_BASE_TAG aligned with the target vLLM release on every rebase"). The ROCm release Dockerfile was already fixed the same way in [ROCm] [CI] Migrate CI to mi300x for v0.27.x #5886.

Evidence

• Image label inspection of the published vllm/vllm-omni:v0.27.0rc1 via the Docker Hub registry API shows ai.vllm.image.tag = vllm/vllm-openai:v0.26.0. Labels are inherited from the actual build base; Dockerfile.cuda has no LABEL overrides and does not reinstall vLLM.
• Timeline: upstream vllm/vllm-openai:v0.27.0 was published 2026-08-10, ~20h before the v0.27.0rc1 image build (2026-08-11), so a correct base was available.
• The release pipeline builds with -f docker/Dockerfile.cuda and passes no --build-arg BASE_IMAGE override, so the stale default is what ships.
• Prior attempt: [Bugfix] Build CUDA images on vLLM 0.27 #6378 (same fix, closed unmerged by its author; no maintainer objection on record).

Test plan

• CI builds the release image with the v0.27.0 base
• Maintainers may want to rebuild/republish the nightly CUDA image after merge so it stops shipping the mismatched runtime; the v0.27.0rc1 tag remains affected (immutable)

[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.

[vllm-omni-review-bot]

This PR was classified as CI work.

CI owner: @yenuo26

@panpan0000, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[panpan0000]

Self-review: one-line change, ARG BASE_IMAGE=vllm/vllm-openai:v0.26.0 → v0.27.0 in docker/Dockerfile.cuda.

What I checked:

• docker/Dockerfile.cuda is a 22-line file with no LABEL overrides and no vLLM reinstall step, so the image's vLLM version is entirely determined by the base image.
• .buildkite/release/release-pipeline.yml builds this Dockerfile (x86_64 and aarch64) with no --build-arg BASE_IMAGE override, so the default is what ships in release and nightly CUDA images.
• docker/Dockerfile.ci already targets VLLM_BASE_TAG=v0.27.0 (kept aligned per its own rebase comment), and docker/Dockerfile.rocm was bumped to vllm-openai-rocm:v0.27.0 in [ROCm] [CI] Migrate CI to mi300x for v0.27.x #5886 — this change makes the CUDA release image consistent with both.
• The published vllm/vllm-omni:v0.27.0rc1 image's inherited label ai.vllm.image.tag=vllm/vllm-openai:v0.26.0 confirms it was built on the 0.26.0 base. Upstream vllm/vllm-openai:v0.27.0 was published 2026-08-10, before the rc1 image build.
• Grepped docker/ and .buildkite/ on main: this line is the only remaining vllm-openai:v0.26.0 pin.

Not verified: I did not run a full image build or boot test locally (no GPU host); relying on CI for that. The already-published v0.27.0rc1 tag stays as-is (immutable); this fix takes effect for nightlies and the v0.27.0 final.

@vllm-omni-review-bot automated review welcome.

[panpan0000]

@lengrongfu PTAL

[hsliuustc0106]

Critical version alignment fix. The CUDA release images were built on vLLM 0.26.0 base while codebase uses vLLM 0.27.0 APIs - this prevents runtime compatibility issues. Simple but essential change.

[panpan0000]

Thank you @hsliuustc0106

[panpan0000]

hmm.. my pr was late..

[NickCao]

hmm.. my pr was late..

Yeah, no longer applicable as we are now on 0.28:

vllm-omni/docker/Dockerfile.cuda

Line 1 in ea14b04

ARG BASE_IMAGE=vllm/vllm-openai:v0.28.0