refactor: install vLLM ROCm from upstream wheels

[dmedovich]

Replaces the build-from-source approach (introduced in #917) with the upstream vLLM ROCm wheels published at https://wheels.vllm.ai/rocm/, as suggested by @ericcurtin.

Builds now take ~3-5 min instead of ~30-60 min, and the runtime image keeps the llamacpp ROCm base — restoring llama.cpp as an in-image fallback (matching the shape of the CUDA vllm variant).

vLLM ROCm wheels follow a separate release cadence from CUDA, so VLLM_ROCM_VERSION (default 0.21.0) is tracked independently from VLLM_VERSION (CUDA, default 0.19.1).

Refs #412

[sourcery-ai]

Hey - I've found 1 issue, and left some high level feedback:

• Consider explicitly pinning the vLLM package to VLLM_ROCM_VERSION in the uv install step (e.g., vllm==${VLLM_ROCM_VERSION}) so the wheel index version and the installed package version cannot drift apart.
• In the apt install step, you can add --no-install-recommends to reduce the image size and avoid pulling in unnecessary packages.

Prompt for AI Agents

Please address the comments from this code review:

## Overall Comments
- Consider explicitly pinning the vLLM package to VLLM_ROCM_VERSION in the uv install step (e.g., `vllm==${VLLM_ROCM_VERSION}`) so the wheel index version and the installed package version cannot drift apart.
- In the `apt install` step, you can add `--no-install-recommends` to reduce the image size and avoid pulling in unnecessary packages.

## Individual Comments

### Comment 1
<location path="Dockerfile" line_range="183-184" />
<code_context>
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && ~/.local/bin/uv venv --python 3.12 /opt/vllm-env \
+    && . /opt/vllm-env/bin/activate \
+    && ~/.local/bin/uv pip install vllm \
+         --extra-index-url https://wheels.vllm.ai/rocm/${VLLM_ROCM_VERSION}/${VLLM_ROCM_TARGET} \
+         --index-strategy unsafe-best-match

</code_context>
<issue_to_address>
**issue (bug_risk):** Tie the installed vLLM package version to VLLM_ROCM_VERSION to avoid pulling a mismatched wheel from PyPI.

Because `uv pip install vllm` resolves against PyPI first, it may select a newer vLLM without ROCm wheels and then fall back to non-ROCm wheels, silently breaking the ROCm build and reproducibility. Please pin the version to match the ROCm wheel path, e.g.:

```dockerfile
&& ~/.local/bin/uv pip install "vllm==${VLLM_ROCM_VERSION}" \
     --extra-index-url https://wheels.vllm.ai/rocm/${VLLM_ROCM_VERSION}/${VLLM_ROCM_TARGET} \
     --index-strategy unsafe-best-match
```

so both the version and wheel source are driven by `VLLM_ROCM_VERSION`.
</issue_to_address>

---

Sourcery is free for open source - if you like our reviews please consider sharing them ✨

• X
• Mastodon
• LinkedIn
• Facebook

_{Help me be more useful! Please click 👍 or 👎 on each comment and I'll use the feedback to improve your reviews.}

[gemini-code-assist]

Code Review

Review: vLLM ROCm Wheel Integration\n\n### Summary\nThis PR refactors the vLLM ROCm build process to use pre-built wheels from the official repository instead of building from source, which significantly reduces build time and aligns the architecture with the CUDA variant.\n\n### Critical\n- Dockerfile:184: The --extra-index-url is likely malformed as the repository is typically flat and the ROCm target version (rocm722) appears to be a typo. Fix: --extra-index-url https://wheels.vllm.ai/rocm

[gemini-code-assist]

The extra index URL appears to be malformed. The official vLLM ROCm wheel repository at wheels.vllm.ai/rocm is a flat repository and typically does not use versioned subdirectories in the path. Furthermore, versioned paths in the vLLM project usually require a v prefix (e.g., v0.21.0). Additionally, rocm722 is likely a typo for a current ROCm version like rocm61 or rocm62. Using an invalid URL will cause the build to fail.

         --extra-index-url https://wheels.vllm.ai/rocm \

[ericcurtin]

@dmedovich sometimes the bots are wrong, but can you confirm if they are right or wrong?

[dmedovich]

@ericcurtin happy to confirm:

Gemini's review — wrong on all three points:

1. "wheels.vllm.ai/rocm is a flat repo without versioned subdirs" — false. Verified:

   $ curl -sI https://wheels.vllm.ai/rocm/0.21.0/rocm722/
   HTTP/2 200
   

   The index returns vllm-0.21.0+rocm722-cp312-cp312-manylinux_2_34_x86_64.whl.

2. "Needs v prefix (v0.21.0)" — false. The existing CUDA stage in the same Dockerfile uses wheels.vllm.ai/${VLLM_VERSION}/cu130 without a v prefix and works fine.

3. "rocm722 is a typo, should be rocm61/rocm62" — false. rocm722 is exactly the suffix you suggested in your comment on feat: add vLLM ROCm Docker variant for AMD GPUs #917 (https://wheels.vllm.ai/rocm/0.21.0/rocm722), which this PR implements.

Sourcery's review — correct, fixed in last commit (2dbc908):

Without pinning vllm==${VLLM_ROCM_VERSION}, uv resolves against both PyPI and our extra index. If a newer vLLM (say 0.22.0) lands on PyPI before the matching ROCm wheel is published, --index-strategy unsafe-best-match would pick the higher PyPI version — silently installing a CUDA-only wheel into the ROCm image. Now pinned.