Building the Images

The blivioniag/rocm-rdna and blivioniag/vllm-rdna images are produced by the vllm-rdna-docker repo. It’s deliberately small: two Dockerfiles, one bake graph, one CI workflow — no Python layer, no custom linter.

Layout

File	Purpose
`Dockerfile.base`	The ROCm + PyTorch + Triton base (`rocm-rdna`).
`Dockerfile.vllm`	Clones and builds vLLM on top of a base image (`vllm-rdna`).
`docker-bake.hcl`	Source of truth — defines every target, tag, and build arg.
`patches/*.patch`	RDNA-specific fixes applied after the vLLM clone.
`.github/workflows/build.yml`	CI: builds on tag push (`v*`) and manual dispatch.

The seven RDNA archs gfx1030;gfx1100;gfx1101;gfx1150;gfx1151;gfx1200;gfx1201 are baked into every image via PYTORCH_ROCM_ARCH.

Build locally

Everything runs through stock docker buildx bake:

# Build everything (bases + vLLM images)
docker buildx bake --file docker-bake.hcl all

# Just the base images, or just the vLLM images
docker buildx bake --file docker-bake.hcl all-bases
docker buildx bake --file docker-bake.hcl all-vllm

# A single target
docker buildx bake --file docker-bake.hcl vllm-0271-rocm720

# Print the plan without building
docker buildx bake --file docker-bake.hcl --print all

Targets are named vllm-<source>-<base>, e.g. vllm-0271-rocm720, vllm-0271-rocm720-extras. Groups: all, all-bases, all-vllm.

Key build arguments (`Dockerfile.vllm`)

Set per-target in docker-bake.hcl:

ARG	Purpose
`BASE_IMAGE`	Published base image, e.g. `blivioniag/rocm-rdna:7.2.0`.
`VLLM_REPOSITORY` / `VLLM_REF`	vLLM clone URL and git ref (tag or branch). Official extras source is `opengfx1030/vllm-rdna`; current bake still clones `https://github.com/BlivionIaG/vllm.git` @ `rdna2_extras`.
`VLLM_COMMIT`	Full 40-char commit; the build fails if `HEAD` doesn’t match (reproducibility).
`VLLM_VARIANT`	`upstream` or `extras-fork` (recorded as an image label).
`PYTORCH_ROCM_ARCH`	Semicolon-joined gfx targets.
`TORCH_BACKEND`	`uv --torch-backend` value, e.g. `rocm7.2`.
`FLASH_ATTENTION_INSTALL`	`base` \| `vllm` \| `none`, plus `FLASH_ATTENTION_REPO` / `_REF`.
`USE_SCCACHE`	`1` to wrap HIP compilation in sccache (base must also be built with it).
`VLLM_PATCH_FILE`	A `.patch` in `patches/` applied after the clone (empty = none).

The `patches/` directory

vLLM occasionally needs small RDNA fixes (e.g. platform detection so consumer Radeon cards are picked up). The repo keeps these as versioned diffs, for example:

patches/v0.26.0-rocm-platforms.patch
patches/v0.27.1-rocm-platform-detect.patch
patches/v0.27.1-amdsmi-wrapper-guard.patch
patches/v0.27.1-extras-*.patch

To add one for a new vLLM release: reproduce the fix against a fresh clone, git diff the changed file(s) into patches/, verify it applies against a clean checkout of that tag, then point the target’s VLLM_PATCH_FILE at it.

Adding a new base or vLLM source

New base: copy a target "base-<id>" block, set the ROCm/PyTorch/Triton versions and index URL, and add the id to all-bases + all-vllm.
New vLLM source: add a target "vllm-<source>-<base>" block per base with VLLM_REPOSITORY, VLLM_REF, VLLM_COMMIT, VLLM_VARIANT, IMAGE_TAG, and add the ids to all-vllm.

See the vllm-rdna-docker README for the full, authoritative instructions.

CI

.github/workflows/build.yml builds and pushes to docker.io/blivioniag/ on tag push (v*), and supports manual dispatch of a single target (optionally pushing). It uses stock docker/setup-buildx, docker/login, docker/metadata, and docker/bake actions with GitHub Actions cache.

Keyboard shortcuts

Wiki GFX1030