[alec-flowers]

Purpose

Installing a vLLM wheel currently places vllm-rs inside the Python package, so the native command cannot be invoked by name. Package it with RustBin so installers place it in the environment's scripts directory, including /usr/local/bin in the standard container and bin in virtual environments.

Draft alternative to #57606, which adds a symlink only to the CUDA serving image. This changes wheel packaging for all installers so vllm-rs directly executes the native binary. The existing vllm serve Python launcher, including VLLM_USE_RUST_FRONTEND=1 vllm serve, remains supported. The container's default vllm serve entrypoint is unchanged. These are alternative approaches to compare, not changes intended to merge together. Searched open PRs for RustBin, "vllm-rs" "packaging", and "vllm-rs" "console"; no other PR implements this packaging change. #51503 concerns a separate standalone renderer image.

Implementation

• Declare the frontend as RustBin. Keep build_rust.sh output at vllm/vllm-rs by selecting that scripts build directory, preserving existing Docker staging paths.
• When native artifacts already exist, copy the frontend into the scripts build directory instead of packaging a second copy inside vllm.
• Accept both vllm/vllm-rs and *.data/scripts/vllm-rs when extracting precompiled wheels.
• Preserve source-local binary discovery, then inspect installed distribution records for the command. Skip source egg-info without an executable so native editable installs work from the checkout. No PATH search is needed.

Four files changed; no Dockerfile or documentation changes. AI assistance was used to prepare and review this prototype.

Validation

Passed on Linux x86_64:

.venv/bin/python -m pytest --noconftest tests/test_envs.py -k 'rust or precompiled' -q
.venv/bin/pre-commit run --files setup.py tools/build_rust.py vllm/envs.py tests/test_envs.py
git diff --check

The focused pytest run passed all four cases. The new parameterized regression covers executable discovery with source metadata shadowing the installed distribution, and preference for an in-place source build.

Packaging checks used the actual setup.py with VLLM_TARGET_DEVICE=empty, prebuilt Rust artifacts, and CARGO set to a nonexistent path:

• Built wheels contain one *.data/scripts/vllm-rs and retain the Rust parser extension.
• uv pip install --no-deps installs an ELF executable; vllm-rs --help and automatic discovery pass. Uninstall removes the command.
• uv pip install --no-deps --no-build-isolation -e . with VLLM_PRECOMPILED_WHEEL_LOCATION succeeds using both old and new wheel layouts. Discovery also passes from the checkout after removing the source-local binary.
• Repeated with actual native artifacts from the pinned release (v0.29.0, image digest sha256:082ca6f035279109041ffd3fe0695cb568b29bc580b35c4f297a66a08b216c1b) and nightly (0.29.1rc1.dev347+gdee37d891, image digest sha256:b88541ab10490f6ee43d2e352edcfcb4b4e0e6b1fdd26ed427c89d73e3fa6a7b).
• Installed the prototype wheels into derived images from both pinned containers with uv pip install --system --no-deps --no-index. /usr/local/bin/vllm-rs is a native executable, not a symlink. docker run --entrypoint vllm-rs ... --help passed as root and as UID 2000.
• A minimal dependency-free Cargo fixture reused the actual frontend declaration and standalone build function to verify source-built wheels, the unchanged staging location, and native editable installation. Passed with setuptools-rust 1.9.0 and 1.13.0; precompiled packaging checks also passed with both versions.

Full vLLM Rust/CUDA compilation and model inference were not run. The Cargo fixture checks packaging destinations, not full frontend compilation. Model evaluation is not applicable to this packaging-only change.