c89d692cb5 · tduncan/Nvidia-Optimization-Stack-VLLM

.buildkite/scripts/ci-bake-rocm.sh

+4 -4

@@ -23,8 +23,8 @@ DEFAULT_CI_BASE_METADATA_VERSION="3"

23 # local-source stages rather than unreachable remote-fetch alternatives.

24 DEFAULT_ROCM_CSRC_CONTENT_FILES=".dockerignore requirements/common.txt requirements/rocm.txt pyproject.toml setup.py CMakeLists.txt cmake csrc vllm/envs.py vllm/__init__.py tools/build_rust.py"

25 DEFAULT_ROCM_CSRC_DOCKERFILE_STAGES="base fetch_vllm_0 fetch_vllm build_vllm_dependencies rocm-triton-kernels csrc-build"

26 -DEFAULT_ROCM_RUST_CONTENT_FILES=".dockerignore requirements/build/rust.txt rust/Cargo.lock rust/Cargo.toml rust/proto rust/src rust-toolchain.toml tools/build_rust.py tools/install_protoc.sh build_rust.sh"

27 -DEFAULT_ROCM_RUST_DOCKERFILE_STAGES="base rust_toolchain_input_0 rust-toolchain-input rust_input_0 rust-input rust-toolchain rust-build"

26 +DEFAULT_ROCM_RUST_CONTENT_FILES=".dockerignore .git_archival.txt pyproject.toml requirements/build/rust.txt rust/Cargo.lock rust/Cargo.toml rust/proto rust/src rust-toolchain.toml tools/build_rust.py tools/install_protoc.sh build_rust.sh"

27 +DEFAULT_ROCM_RUST_DOCKERFILE_STAGES="base fetch_vllm_0 fetch_vllm vllm-version rust_toolchain_input_0 rust-toolchain-input rust_input_0 rust-input rust-toolchain rust-build"

28 # Docker's 128-character tag limit minus the longest cache prefix

29 # ("csrc-rocm-branch-" and "rust-rocm-branch-", both 17 characters).

30 ROCM_CACHE_BRANCH_TAG_MAX_LEN=111

@@ -523,8 +523,8 @@ prepare_ci_build_context() {

523 return 1

524 fi

525

526 - # setuptools-scm understands Git's stable archive format, so full-source

527 - # wheels retain their exact version without copying mutable Git history.

526 + # setuptools-scm understands Git's stable archive format, so wheels and

527 + # Rust artifacts retain their exact version without copying Git history.

528 if ! is_ci_base_target; then

529 write_ci_git_archival_metadata "${source_root}" "${context_root}" \

530 || return $?

.buildkite/test-amd.yaml

+1 -1

@@ -2356,7 +2356,7 @@ steps:

2356 - export VLLM_USE_RUST_FRONTEND=1

2357 - export VLLM_WORKER_MULTIPROC_METHOD=spawn

2358 - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc/test_collective_rpc.py

2359 - - pytest -v -s entrypoints/serve/instrumentator/test_basic.py -k "not show_version and not server_load"

2359 + - pytest -v -s entrypoints/serve/instrumentator/test_basic.py -k "not server_load"

2360 - pytest -v -s entrypoints/scale_out/token_in_token_out/test_serving_tokens.py -k "not stream and not lora and not test_generate_logprobs and not stop_string_workflow"

2361 - pytest -v -s entrypoints/serve/instrumentator/test_metrics.py -k "text and not show and not run_batch and not test_metrics_counts and not test_metrics_exist"

2362 - pytest -v -s entrypoints/serve/tokenize/test_tokenization.py -k "not tokenizer_info"

.buildkite/test_areas/rust_frontend.yaml

+1 -1

@@ -58,7 +58,7 @@ steps:

58 - export VLLM_WORKER_MULTIPROC_METHOD=spawn

59 - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc/test_collective_rpc.py

60 # server_load can be flaky under the Rust frontend; keep it excluded for now.

61 - - pytest -v -s entrypoints/serve/instrumentator/test_basic.py -k "not show_version and not server_load"

61 + - pytest -v -s entrypoints/serve/instrumentator/test_basic.py -k "not server_load"

62 # test_generate_logprobs expects Python-style top_logprobs truncation (dedup sampled + cap at max(k, 1)).

63 - pytest -v -s entrypoints/scale_out/token_in_token_out/test_serving_tokens.py -k "not stream and not lora and not test_generate_logprobs and not stop_string_workflow"

64 - pytest -v -s entrypoints/serve/instrumentator/test_metrics.py -k "text and not show and not run_batch and not test_metrics_counts and not test_metrics_exist"

docker/Dockerfile

+5

@@ -279,11 +279,16 @@ ENV CARGO_BUILD_JOBS=4

279 # separate local sccache daemon while sharing the same remote cache backend.

280 ENV SCCACHE_SERVER_PORT=4227

281

282 +# Only Rust inputs are present, so suppress the artificial dirty state caused

283 +# by other tracked files being absent from this build stage.

284 +ENV SETUPTOOLS_SCM_PRETEND_METADATA={dirty=false}

285 +

286 # Build the release artifacts. Cache cargo registry/git, but not target/,

287 # because stale target metadata can outlive source updates across BuildKit

288 # cache reuse.

289 RUN --mount=type=cache,target=/root/.cargo/registry,sharing=locked \

290 --mount=type=cache,target=/root/.cargo/git,sharing=locked \

291 + --mount=type=bind,source=.git,target=.git \

292 --mount=type=secret,id=aws-credentials,target=/root/.aws/credentials,required=false \

293 if [ "$USE_SCCACHE" = "1" ]; then \

294 if [ -n "${SCCACHE_ENDPOINT}" ]; then export SCCACHE_ENDPOINT="${SCCACHE_ENDPOINT}"; fi; \

docker/Dockerfile.cpu

+5

@@ -166,11 +166,16 @@ COPY build_rust.sh build_rust.sh

166 # (rustc spawns enough concurrent processes to hit RLIMIT_NOFILE otherwise).

167 ENV CARGO_BUILD_JOBS=4

168

169 +# Only Rust inputs are present, so suppress the artificial dirty state caused

170 +# by other tracked files being absent from this build stage.

171 +ENV SETUPTOOLS_SCM_PRETEND_METADATA={dirty=false}

172 +

173 # Build the release artifacts. Cache cargo registry/git, but not target/,

174 # because stale target metadata can outlive source updates across BuildKit

175 # cache reuse.

176 RUN --mount=type=cache,target=/root/.cargo/registry,sharing=locked \

177 --mount=type=cache,target=/root/.cargo/git,sharing=locked \

178 + --mount=type=bind,source=.git,target=.git \

179 --mount=type=secret,id=aws-credentials,target=/root/.aws/credentials,required=false \

180 if [ "$USE_SCCACHE" = "1" ]; then \

181 export RUSTC_WRAPPER=sccache; \

docker/Dockerfile.rocm

+23 -12

@@ -125,6 +125,18 @@ ONBUILD RUN git clone ${VLLM_REPO} \

125 && git fetch upstream ; fi

126 FROM fetch_vllm_${REMOTE_VLLM} AS fetch_vllm

127

128 +# Resolve the device-independent source version once for every downstream

129 +# artifact. setuptools-scm reads Git metadata for local/remote sources and the

130 +# archive metadata generated for canonical CI contexts.

131 +FROM fetch_vllm AS vllm-version

132 +ARG COMMON_WORKDIR

133 +RUN --mount=type=cache,id=vllm-rocm-uv,target=/root/.cache/uv \

134 + cd ${COMMON_WORKDIR}/vllm \

135 + && uv pip install --system -r requirements/build/rust.txt \

136 + && python3 -c \

137 + "from setuptools_scm import get_version; print(get_version())" \

138 + > /vllm-version.txt

139 +

140 # -----------------------

141 # Rust/protoc toolchain inputs

142 #

@@ -238,6 +250,7 @@ COPY --from=rust-input /rust-input/vllm/rust rust

250 COPY --from=rust-input /rust-input/vllm/rust-toolchain.toml rust-toolchain.toml

251 COPY --from=rust-input /rust-input/vllm/tools/build_rust.py tools/build_rust.py

252 COPY --from=rust-input /rust-input/vllm/build_rust.sh build_rust.sh

253 +COPY --from=vllm-version /vllm-version.txt /tmp/vllm-version.txt

254

255 # Build the release binary. Cargo's registry/git caches can be written by

256 # concurrent BuildKit jobs on shared workers, so lock those cache mounts while

@@ -249,6 +262,7 @@ RUN --mount=type=cache,id=vllm-rocm-cargo-registry,target=/root/.cargo/registry,

262 export RUSTC_WRAPPER=sccache \

263 && sccache --show-stats; \

264 fi \

265 + && export SETUPTOOLS_SCM_PRETEND_VERSION="$(cat /tmp/vllm-version.txt)" \

266 && bash build_rust.sh \

267 && test -x vllm/vllm-rs \

268 && if [ "$USE_SCCACHE" = "1" ]; then \

@@ -334,6 +348,7 @@ RUN --mount=type=bind,source=pyproject.toml,target=${COMMON_WORKDIR}/vllm/pyproj

348 FROM build_vllm_dependencies AS build_vllm

349 COPY --from=fetch_vllm ${COMMON_WORKDIR}/vllm ${COMMON_WORKDIR}/vllm

350 COPY --from=csrc-build ${COMMON_WORKDIR}/vllm/dist /precompiled-wheels

351 +COPY --from=vllm-version /vllm-version.txt /tmp/vllm-version.txt

352

353 # Drop the pre-built Rust artifacts into the source tree. setup.py detects

354 # them and ships them as-is, skipping the local Rust build.

@@ -341,6 +356,7 @@ COPY --from=rust-build ${COMMON_WORKDIR}/vllm/vllm/vllm-rs ${COMMON_WORKDIR}/vll

356 COPY --from=rust-build ${COMMON_WORKDIR}/vllm/vllm/_rust_*.so ${COMMON_WORKDIR}/vllm/vllm/

357

358 RUN cd "${COMMON_WORKDIR}/vllm" \

359 + && export SETUPTOOLS_SCM_PRETEND_VERSION="$(cat /tmp/vllm-version.txt)" \

360 && export VLLM_USE_PRECOMPILED=1 \

361 && export VLLM_PRECOMPILED_WHEEL_LOCATION="$(ls /precompiled-wheels/*.whl)" \

362 && export VLLM_DOCKER_BUILD_CONTEXT=1 \

@@ -640,6 +656,7 @@ FROM fetch_vllm AS build_vllm_wheel_release

656

657 ARG COMMON_WORKDIR

658 ARG ROCM_TRITON_KERNELS_COMMIT

659 +COPY --from=vllm-version /vllm-version.txt /tmp/vllm-version.txt

660 COPY --from=rocm-triton-kernels \

661 /opt/rocm-triton-kernels /opt/rocm-triton-kernels

662 ENV TRITON_KERNELS_SRC_DIR=/opt/rocm-triton-kernels

@@ -678,17 +695,12 @@ RUN if [ "$GIT_REPO_CHECK" != "0" ]; then \

695 cd vllm && bash tools/check_repo.sh; \

696 fi

697

681 -# Extract version from git BEFORE any modifications (pin_rocm_dependencies.py modifies requirements/rocm.txt)

682 -# This ensures setuptools_scm sees clean repo state for version detection

683 -RUN --mount=type=bind,source=.git,target=vllm/.git \

684 - --mount=type=cache,target=/root/.cache/uv \

698 +# Install the packaging helpers before modifying requirements/rocm.txt.

699 +RUN --mount=type=cache,target=/root/.cache/uv \

700 grep -Fq "set(TRITON_KERNELS_TAG \"${ROCM_TRITON_KERNELS_COMMIT}\")" \

701 vllm/cmake/external_projects/triton_kernels.cmake \

702 && cd vllm \

688 - && uv pip install --system setuptools_scm regex \

689 - && VLLM_VERSION=$(python3 -c "import setuptools_scm; print(setuptools_scm.get_version())") \

690 - && echo "Detected vLLM version: ${VLLM_VERSION}" \

691 - && echo "${VLLM_VERSION}" > /tmp/vllm_version.txt

703 + && uv pip install --system setuptools_scm regex

704

705 # Fail if git-based package dependencies are found in requirements files

706 # (uv doesn't handle git+ URLs well, and packages should be distributed on PyPI)

@@ -726,13 +738,12 @@ RUN --mount=type=cache,target=/root/.cache/uv \

738 && echo "Building vLLM with custom wheels from /install" \

739 && uv pip install --system --find-links /install -r requirements/rocm.txt

740

729 -# Build wheel using pre-extracted version to avoid dirty state from modified requirements/rocm.txt

741 +# Build wheel using the shared source version to avoid dirty state from modified requirements/rocm.txt

742 # (setup.py auto-detects ccache/sccache in PATH)

731 -RUN --mount=type=bind,source=.git,target=vllm/.git \

732 - --mount=type=cache,id=vllm-rocm-ccache,target=/root/.cache/ccache \

743 +RUN --mount=type=cache,id=vllm-rocm-ccache,target=/root/.cache/ccache \

744 cd vllm \

745 && export CCACHE_BASEDIR="$PWD" \

735 - && export SETUPTOOLS_SCM_PRETEND_VERSION=$(cat /tmp/vllm_version.txt) \

746 + && export SETUPTOOLS_SCM_PRETEND_VERSION="$(cat /tmp/vllm-version.txt)" \

747 && echo "Building wheel with version: ${SETUPTOOLS_SCM_PRETEND_VERSION}" \

748 && MAX_JOBS="${MAX_JOBS:-$(nproc)}" python3 setup.py bdist_wheel --dist-dir=dist

749

docker/Dockerfile.rocm_gfx1250

+23 -12

@@ -127,6 +127,17 @@ ONBUILD RUN git clone ${VLLM_REPO} \

127 && git fetch upstream ; fi

128 FROM fetch_vllm_${REMOTE_VLLM} AS fetch_vllm

129

130 +# Resolve the device-independent source version once for every downstream

131 +# artifact. setuptools-scm reads Git metadata from local and remote sources.

132 +FROM fetch_vllm AS vllm-version

133 +ARG COMMON_WORKDIR

134 +RUN --mount=type=cache,id=vllm-rocm-uv,target=/root/.cache/uv \

135 + cd ${COMMON_WORKDIR}/vllm \

136 + && uv pip install --system -r requirements/build/rust.txt \

137 + && python3 -c \

138 + "from setuptools_scm import get_version; print(get_version())" \

139 + > /vllm-version.txt

140 +

141 # -----------------------

142 # Rust build stage

143 # Builds the `vllm-rs` frontend in a dedicated stage so the wheel build stages

@@ -152,6 +163,8 @@ RUN --mount=type=cache,id=vllm-rocm-uv,target=/root/.cache/uv \

163 cd ${COMMON_WORKDIR}/vllm \

164 && uv pip install --system -r requirements/build/rust.txt

165

166 +COPY --from=vllm-version /vllm-version.txt /tmp/vllm-version.txt

167 +

168 # Build the release binary. Cargo's registry/git caches can be written by

169 # concurrent BuildKit jobs on shared workers, so lock those cache mounts while

170 # keeping the cache benefit. Do not cache target/, because stale target metadata

@@ -159,6 +172,7 @@ RUN --mount=type=cache,id=vllm-rocm-uv,target=/root/.cache/uv \

172 RUN --mount=type=cache,id=vllm-rocm-cargo-registry,target=/root/.cargo/registry,sharing=locked \

173 --mount=type=cache,id=vllm-rocm-cargo-git,target=/root/.cargo/git,sharing=locked \

174 cd ${COMMON_WORKDIR}/vllm \

175 + && export SETUPTOOLS_SCM_PRETEND_VERSION="$(cat /tmp/vllm-version.txt)" \

176 && bash build_rust.sh \

177 && test -x vllm/vllm-rs

178

@@ -211,6 +225,7 @@ ARG COMMON_WORKDIR

225 ENV VLLM_TARGET_DEVICE=rocm

226

227 COPY --from=csrc-build ${COMMON_WORKDIR}/vllm/dist /precompiled-wheels

228 +COPY --from=vllm-version /vllm-version.txt /tmp/vllm-version.txt

229

230 # Drop the pre-built Rust artifacts into the source tree. setup.py detects

231 # them and ships them as-is, skipping the local Rust build.

@@ -220,6 +235,7 @@ COPY --from=rust-build ${COMMON_WORKDIR}/vllm/vllm/_rust_*.so ${COMMON_WORKDIR}/

235 RUN --mount=type=cache,id=vllm-rocm-uv,target=/root/.cache/uv \

236 cd vllm \

237 && uv pip install --system -r requirements/rocm.txt \

238 + && export SETUPTOOLS_SCM_PRETEND_VERSION="$(cat /tmp/vllm-version.txt)" \

239 && export VLLM_USE_PRECOMPILED=1 \

240 && export VLLM_PRECOMPILED_WHEEL_LOCATION="$(ls /precompiled-wheels/*.whl)" \

241 && export VLLM_DOCKER_BUILD_CONTEXT=1 \

@@ -426,6 +442,7 @@ RUN /bin/bash -lc 'set -euo pipefail; \

442 FROM fetch_vllm AS build_vllm_wheel_release

443

444 ARG COMMON_WORKDIR

445 +COPY --from=vllm-version /vllm-version.txt /tmp/vllm-version.txt

446

447 # Drop the pre-built Rust artifacts into the source tree. setup.py detects

448 # them and ships them as-is, skipping the local Rust build.

@@ -461,15 +478,10 @@ RUN if [ "$GIT_REPO_CHECK" != "0" ]; then \

478 cd vllm && bash tools/check_repo.sh; \

479 fi

480

464 -# Extract version from git BEFORE any modifications (pin_rocm_dependencies.py modifies requirements/rocm.txt)

465 -# This ensures setuptools_scm sees clean repo state for version detection

466 -RUN --mount=type=bind,source=.git,target=vllm/.git \

467 - --mount=type=cache,target=/root/.cache/uv \

481 +# Install the packaging helpers before modifying requirements/rocm.txt.

482 +RUN --mount=type=cache,target=/root/.cache/uv \

483 cd vllm \

469 - && uv pip install --system setuptools_scm regex \

470 - && VLLM_VERSION=$(python3 -c "import setuptools_scm; print(setuptools_scm.get_version())") \

471 - && echo "Detected vLLM version: ${VLLM_VERSION}" \

472 - && echo "${VLLM_VERSION}" > /tmp/vllm_version.txt

484 + && uv pip install --system setuptools_scm regex

485

486 # Fail if git-based package dependencies are found in requirements files

487 # (uv doesn't handle git+ URLs well, and packages should be distributed on PyPI)

@@ -507,13 +519,12 @@ RUN --mount=type=cache,target=/root/.cache/uv \

519 && echo "Building vLLM with custom wheels from /install" \

520 && uv pip install --system --find-links /install -r requirements/rocm.txt

521

510 -# Build wheel using pre-extracted version to avoid dirty state from modified requirements/rocm.txt

522 +# Build wheel using the shared source version to avoid dirty state from modified requirements/rocm.txt

523 # (setup.py auto-detects ccache/sccache in PATH)

512 -RUN --mount=type=bind,source=.git,target=vllm/.git \

513 - --mount=type=cache,id=vllm-rocm-ccache,target=/root/.cache/ccache \

524 +RUN --mount=type=cache,id=vllm-rocm-ccache,target=/root/.cache/ccache \

525 cd vllm \

526 && export CCACHE_BASEDIR="$PWD" \

516 - && export SETUPTOOLS_SCM_PRETEND_VERSION=$(cat /tmp/vllm_version.txt) \

527 + && export SETUPTOOLS_SCM_PRETEND_VERSION="$(cat /tmp/vllm-version.txt)" \

528 && echo "Building wheel with version: ${SETUPTOOLS_SCM_PRETEND_VERSION}" \

529 && MAX_JOBS="${MAX_JOBS:-$(nproc)}" python3 setup.py bdist_wheel --dist-dir=dist

530

docker/Dockerfile.xpu

+5

@@ -28,8 +28,13 @@ COPY build_rust.sh build_rust.sh

28 # (rustc spawns enough concurrent processes to hit RLIMIT_NOFILE otherwise).

29 ENV CARGO_BUILD_JOBS=4

30

31 +# Only Rust inputs are present, so suppress the artificial dirty state caused

32 +# by other tracked files being absent from this build stage.

33 +ENV SETUPTOOLS_SCM_PRETEND_METADATA={dirty=false}

34 +

35 RUN --mount=type=cache,target=/root/.cargo/registry,sharing=locked \

36 --mount=type=cache,target=/root/.cargo/git,sharing=locked \

37 + --mount=type=bind,source=.git,target=.git \

38 bash build_rust.sh

39

40 FROM ubuntu:24.04 AS vllm-base

requirements/build/rust.txt

+1

@@ -1,4 +1,5 @@

1 # Dependencies for building Rust artifacts through setuptools-rust.

2 setuptools>=77.0.3,<81.0.0

3 +setuptools-scm>=9.2.0

4 setuptools-rust>=1.9.0

5 wheel

rust/Cargo.lock

+7

@@ -5505,9 +5505,14 @@ dependencies = [

5505 "tracing",

5506 "url",

5507 "uuid",

5508 + "vllm-build-info",

5509 "vllm-tracing",

5510 ]

5511

5512 +[[package]]

5513 +name = "vllm-build-info"

5514 +version = "0.1.0"

5515 +

5516 [[package]]

5517 name = "vllm-chat"

5518 version = "0.1.0"

@@ -5574,6 +5579,7 @@ dependencies = [

5579 "tracing",

5580 "uuid",

5581 "vllm-bench",

5582 + "vllm-build-info",

5583 "vllm-chat",

5584 "vllm-engine-core-client",

5585 "vllm-managed-engine",

@@ -5755,6 +5761,7 @@ dependencies = [

5761 "url",

5762 "uuid",

5763 "validator",

5764 + "vllm-build-info",

5765 "vllm-chat",

5766 "vllm-engine-core-client",

5767 "vllm-llm",

rust/Cargo.toml

+2

@@ -1,6 +1,7 @@

1 [workspace]

2 members = [

3 "src/bench",

4 + "src/build-info",

5 "src/chat",

6 "src/cmd",

7 "src/engine-core-client",

@@ -135,6 +136,7 @@ url = "2.5.7"

136 uuid = { version = "1.22.0", features = ["v4"] }

137 validator = { version = "0.20.0", features = ["derive"] }

138 vllm-bench = { path = "src/bench" }

139 +vllm-build-info = { path = "src/build-info" }

140 vllm-chat = { path = "src/chat" }

141 vllm-engine-core-client = { path = "src/engine-core-client" }

142 vllm-llm = { path = "src/llm" }

rust/src/bench/Cargo.toml

+1

@@ -34,6 +34,7 @@ tokio-stream.workspace = true

34 tracing.workspace = true

35 url.workspace = true

36 uuid.workspace = true

37 +vllm-build-info.workspace = true

38 vllm-tracing.workspace = true

39

40 [lints]

rust/src/bench/src/main.rs

+1 -1

@@ -12,7 +12,7 @@ use clap::Parser;

12 #[command(

13 name = "vllm-bench",

14 about = "Benchmark online serving throughput",

15 - version

15 + version = vllm_build_info::VERSION

16 )]

17 struct Cli {

18 #[command(flatten)]

rust/src/build-info/Cargo.toml new

+8

@@ -0,0 +1,8 @@

1 +[package]

2 +name = "vllm-build-info"

3 +version.workspace = true

4 +edition.workspace = true

5 +license.workspace = true

6 +

7 +[lints]

8 +workspace = true

rust/src/build-info/src/lib.rs new

+10

@@ -0,0 +1,10 @@

1 +// SPDX-License-Identifier: Apache-2.0

2 +// SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 +

4 +/// The vLLM package version supplied by the build system.

5 +///

6 +/// Direct Cargo builds fall back to the internal crate version.

7 +pub const VERSION: &str = match option_env!("VLLM_RS_BUILD_VERSION") {

8 + Some(version) => version,

9 + None => env!("CARGO_PKG_VERSION"),

10 +};

rust/src/cmd/Cargo.toml

+1

@@ -28,6 +28,7 @@ tokio-util.workspace = true

28 tracing.workspace = true

29 uuid.workspace = true

30 vllm-bench.workspace = true

31 +vllm-build-info.workspace = true

32 vllm-chat.workspace = true

33 vllm-engine-core-client.workspace = true

34 vllm-managed-engine.workspace = true

rust/src/cmd/src/cli.rs

+2 -1

@@ -39,7 +39,8 @@ use crate::cli::unsupported::UnsupportedArgs;

39 #[derive(Debug, Parser)]

40 #[command(

41 name = "vllm-rs",

42 - about = "Rust frontend and managed-engine CLI for vLLM."

42 + about = "Rust frontend and managed-engine CLI for vLLM.",

43 + version = vllm_build_info::VERSION

44 )]

45 pub struct Cli {

46 #[command(subcommand)]

rust/src/server/Cargo.toml

+1

@@ -45,6 +45,7 @@ tracing-subscriber.workspace = true

45 url.workspace = true

46 uuid.workspace = true

47 validator.workspace = true

48 +vllm-build-info.workspace = true

49 vllm-chat.workspace = true

50 vllm-engine-core-client.workspace = true

51 vllm-llm.workspace = true

rust/src/server/src/routes/tests.rs

+1 -1

@@ -1746,7 +1746,7 @@ async fn version_returns_engine_vllm_version() {

1746 json,

1747 json!({

1748 "version": "test-vllm-version",

1749 - "rust_frontend_version": env!("CARGO_PKG_VERSION"),

1749 + "rust_frontend_version": vllm_build_info::VERSION,

1750 })

1751 );

1752 }

rust/src/server/src/routes/version.rs

+1 -1

@@ -21,6 +21,6 @@ pub async fn version(State(state): State<Arc<AppState>>) -> Json<VersionResponse

21

22 Json(VersionResponse {

23 version,

24 - rust_frontend_version: env!("CARGO_PKG_VERSION"),

24 + rust_frontend_version: vllm_build_info::VERSION,

25 })

26 }

setup.py

+6 -1

@@ -1493,6 +1493,11 @@ if (

1493 ):

1494 cmdclass["build_rust"] = precompiled_build_rust

1495

1496 +# Resolve the Python version first because get_vllm_version() may set

1497 +# SETUPTOOLS_SCM_PRETEND_VERSION, which the Rust version should inherit.

1498 +vllm_version = get_vllm_version()

1499 +rust_build.prepare_build_environment()

1500 +

1501 # Rust artifacts, built via setuptools-rust and installed into the package

1502 # directory alongside the Python modules.

1503 rust_extensions = rust_build.rust_extensions(

@@ -1501,7 +1506,7 @@ rust_extensions = rust_build.rust_extensions(

1506

1507 setup(

1508 # static metadata should rather go in pyproject.toml

1504 - version=get_vllm_version(),

1509 + version=vllm_version,

1510 ext_modules=ext_modules,

1511 rust_extensions=rust_extensions,

1512 install_requires=get_requirements(),

tests/entrypoints/openai/test_uds.py

+5 -2

@@ -6,6 +6,7 @@ from tempfile import TemporaryDirectory

6 import httpx

7 import pytest

8

9 +from vllm import envs

10 from vllm.version import __version__ as VLLM_VERSION

11

12 from ...utils import RemoteOpenAIServer

@@ -40,5 +41,7 @@ async def test_show_version(server: RemoteOpenAIServer):

41 response = client.get(server.url_for("version"))

42 response.raise_for_status()

43

43 - # Tolerate additive fields (e.g. the Rust frontend reports its own version).

44 - assert response.json()["version"] == VLLM_VERSION

44 + payload = response.json()

45 + assert payload["version"] == VLLM_VERSION

46 + if envs.VLLM_USE_RUST_FRONTEND:

47 + assert payload["rust_frontend_version"] == VLLM_VERSION

tests/entrypoints/serve/instrumentator/test_basic.py

+5 -2

@@ -12,6 +12,7 @@ import requests

12 from fastapi import Request

13

14 from tests.utils import RemoteOpenAIServer

15 +from vllm import envs

16 from vllm.v1.engine.exceptions import EngineDeadError

17 from vllm.version import __version__ as VLLM_VERSION

18

@@ -83,8 +84,10 @@ async def test_show_version(server: RemoteOpenAIServer):

84 response = requests.get(server.url_for("version"))

85 response.raise_for_status()

86

86 - # Tolerate additive fields (e.g. the Rust frontend reports its own version).

87 - assert response.json()["version"] == VLLM_VERSION

87 + payload = response.json()

88 + assert payload["version"] == VLLM_VERSION

89 + if envs.VLLM_USE_RUST_FRONTEND:

90 + assert payload["rust_frontend_version"] == VLLM_VERSION

91

92

93 @pytest.mark.asyncio

tools/build_rust.py

+16

@@ -11,8 +11,23 @@ from pathlib import Path

11

12 from setuptools import setup

13 from setuptools_rust import Binding, RustExtension

14 +from setuptools_scm import get_version

15

16 ROOT_DIR = Path(__file__).resolve().parents[1]

17 +VLLM_RS_BUILD_VERSION = "VLLM_RS_BUILD_VERSION"

18 +

19 +

20 +def prepare_build_environment() -> str | None:

21 + """Set the device-independent vLLM source version for Rust artifacts."""

22 + version = os.getenv(VLLM_RS_BUILD_VERSION) or None

23 + if version is None:

24 + try:

25 + version = get_version(root=ROOT_DIR)

26 + except LookupError:

27 + return None

28 +

29 + os.environ[VLLM_RS_BUILD_VERSION] = version

30 + return version

31

32

33 def rust_extensions(*, optional: bool = False) -> list[RustExtension]:

@@ -61,6 +76,7 @@ def build_binary(build_rust_args: list[str]) -> None:

76

77

78 def main() -> None:

79 + prepare_build_environment()

80 build_binary(sys.argv[1:])

81

82

[Build] Propagate vLLM version to Rust binaries (#52593)