[CI] Discover Rust coverage artifacts from build metadata

Co-authored-by: OpenAI Codex <codex@openai.com>

.buildkite/scripts/run-rust-frontend-cargo-ci.sh

@@ -29,6 +29,7 @@ PYO3_PYTHON_VERSION="${PYO3_PYTHON_VERSION:-3.12}"
 CARGO_SORT_VERSION_REQ="${CARGO_SORT_VERSION_REQ:-2}"
 CARGO_DENY_VERSION_REQ="${CARGO_DENY_VERSION_REQ:-0.20}"
 CARGO_NEXTEST_VERSION_REQ="${CARGO_NEXTEST_VERSION_REQ:-0.9}"
+CARGO_LLVM_COV_VERSION="${CARGO_LLVM_COV_VERSION:-0.8.7}"
 
 log_section() {
   echo "--- $*"
@@ -106,6 +107,18 @@ install_cargo_nextest() {
     "cargo-nextest@${CARGO_NEXTEST_VERSION_REQ}"
 }
 
+install_cargo_llvm_cov() {
+  log_section "Installing cargo-llvm-cov ${CARGO_LLVM_COV_VERSION}"
+  local toolchain
+  toolchain="$(rust_toolchain)"
+  rustup component add --toolchain "$toolchain" llvm-tools-preview
+  cargo binstall \
+    --no-confirm \
+    --force \
+    --secure \
+    "cargo-llvm-cov@${CARGO_LLVM_COV_VERSION}"
+}
+
 install_uv() {
   log_section "Installing uv ${UV_VERSION}"
   curl -L --proto '=https' --tlsv1.2 -sSf \
@@ -176,14 +189,41 @@ run_tests() {
   setup_pyo3_python
   install_cargo_binstall
   install_cargo_nextest
+  install_cargo_llvm_cov
 
-  log_section "Running cargo nextest"
-  cargo nextest run \
+  log_section "Running cargo nextest with Rust coverage"
+  mkdir -p artifacts
+  export LLVM_PROFILE_FILE_NAME="vllm-rust-unit-%4m.profraw"
+  cargo llvm-cov clean \
+    --manifest-path rust/Cargo.toml \
+    --profraw-only
+
+  set +e
+  cargo llvm-cov nextest \
     --manifest-path rust/Cargo.toml \
     --workspace \
     --all-features \
     --locked \
-    --no-fail-fast
+    --no-fail-fast \
+    --no-clean \
+    --lcov \
+    --output-path artifacts/rust-unit.lcov \
+    --ignore-filename-regex='/\.cargo/(registry|git)/|/rustc/|/target/'
+  local coverage_rc=$?
+
+  local upload_rc=0
+  if [[ $coverage_rc -eq 0 ]]; then
+    # shellcheck source=.buildkite/scripts/rust-coverage.sh
+    source .buildkite/scripts/rust-coverage.sh
+    rust_coverage_upload artifacts/rust-unit.lcov rust-unit
+    upload_rc=$?
+  fi
+  set -e
+
+  if [[ $coverage_rc -ne 0 ]]; then
+    return "$coverage_rc"
+  fi
+  return "$upload_rc"
 }
 
 install_protoc

.buildkite/scripts/rust-coverage.sh

@@ -0,0 +1,182 @@
+#!/bin/sh
+
+RUST_CODECOV_VERSION="v11.3.1"
+RUST_CODECOV_SHA256="ca1d64196d2d34771084afe76ea657d581bf628e31d993ff8e52ea09cc88a56d"
+
+rust_coverage_repo_root() {
+    if [ -f /vllm-workspace/.buildkite/scripts/rust-coverage.sh ]; then
+        printf '%s\n' /vllm-workspace
+    elif [ -n "${BUILDKITE_BUILD_CHECKOUT_PATH:-}" ] \
+        && [ -d "$BUILDKITE_BUILD_CHECKOUT_PATH" ]; then
+        printf '%s\n' "$BUILDKITE_BUILD_CHECKOUT_PATH"
+    else
+        git rev-parse --show-toplevel
+    fi
+}
+
+rust_coverage_start() {
+    RUST_COVERAGE_FLAG=${1:?coverage flag is required}
+    RUST_COVERAGE_DIR="/tmp/vllm-rust-coverage/${BUILDKITE_JOB_ID:-local}"
+    export RUST_COVERAGE_FLAG RUST_COVERAGE_DIR
+    mkdir -p "$RUST_COVERAGE_DIR"
+    LLVM_PROFILE_FILE="$RUST_COVERAGE_DIR/rust-%4m.profraw"
+    export LLVM_PROFILE_FILE
+    trap rust_coverage_finalize 0
+}
+
+rust_coverage_objects() {
+    rust_cov_objects_manifest="$(dirname "$(command -v llvm-cov)")/../objects"
+    python3 - "$rust_cov_objects_manifest" <<'PY'
+from pathlib import Path
+import sys
+
+for relative in Path(sys.argv[1]).read_text().splitlines():
+    for entry in sys.path:
+        path = Path(entry or ".").resolve() / relative
+        if path.is_file():
+            print(path)
+            break
+    else:
+        raise RuntimeError(f"installed Rust coverage object was not found: {relative}")
+PY
+}
+
+rust_coverage_collect() {
+    rust_cov_collect_flag=${1:?coverage flag is required}
+    rust_cov_collect_lcov="$RUST_COVERAGE_DIR/$rust_cov_collect_flag.lcov"
+
+    rust_cov_collect_objects=$(rust_coverage_objects) || return 1
+    rust_cov_collect_primary=
+    set --
+    while IFS= read -r rust_cov_collect_object; do
+        if [ -z "$rust_cov_collect_primary" ]; then
+            rust_cov_collect_primary=$rust_cov_collect_object
+        else
+            set -- "$@" "--object=$rust_cov_collect_object"
+        fi
+    done <<EOF
+$rust_cov_collect_objects
+EOF
+
+    llvm-profdata merge \
+        -sparse \
+        "$RUST_COVERAGE_DIR"/*.profraw \
+        -o "$RUST_COVERAGE_DIR/merged.profdata" || return 1
+    llvm-cov export \
+        "$rust_cov_collect_primary" \
+        "$@" \
+        --format=lcov \
+        --instr-profile="$RUST_COVERAGE_DIR/merged.profdata" \
+        --ignore-filename-regex='/\.cargo/(registry|git)/|/rustc/|/target/' \
+        > "$rust_cov_collect_lcov" || return 1
+    RUST_COVERAGE_LCOV=$rust_cov_collect_lcov
+    export RUST_COVERAGE_LCOV
+}
+
+rust_coverage_upload() {
+    rust_cov_upload_lcov=${1:?LCOV path is required}
+    rust_cov_upload_flag=${2:?coverage flag is required}
+    rust_cov_upload_repo_root=$(rust_coverage_repo_root) || return 1
+
+    if [ "$(uname -m)" != "x86_64" ]; then
+        echo "Rust coverage upload currently supports x86_64 CI agents" >&2
+        return 1
+    fi
+
+    rust_cov_upload_codecov_dir=$(mktemp -d /tmp/codecov-bin.XXXXXX) \
+        || return 1
+    curl -fsSL \
+        "https://github.com/codecov/codecov-cli/releases/download/${RUST_CODECOV_VERSION}/codecovcli_linux" \
+        -o "$rust_cov_upload_codecov_dir/codecov" || return 1
+    echo "$RUST_CODECOV_SHA256  $rust_cov_upload_codecov_dir/codecov" \
+        | sha256sum -c - || return 1
+    chmod +x "$rust_cov_upload_codecov_dir/codecov" || return 1
+
+    rust_cov_upload_slug="vllm-project/vllm"
+    if [ -n "${BUILDKITE_PULL_REQUEST:-}" ] \
+        && [ "${BUILDKITE_PULL_REQUEST}" != "false" ] \
+        && [ -n "${BUILDKITE_PULL_REQUEST_REPO:-}" ]; then
+        rust_cov_upload_slug=$(echo "$BUILDKITE_PULL_REQUEST_REPO" \
+            | sed -E 's#(git@|https?://)([^/:]+)[:/]([^/]+/[^/.]+)(\.git)?$#\3#')
+        case "$rust_cov_upload_slug" in
+            */*) ;;
+            *) rust_cov_upload_slug="vllm-project/vllm" ;;
+        esac
+    fi
+
+    rust_cov_upload_branch=${BUILDKITE_BRANCH:?BUILDKITE_BRANCH is required}
+    if [ -z "${CODECOV_TOKEN:-}" ]; then
+        # Codecov accepts tokenless public uploads on unprotected branch names.
+        # A colon-separated prefix keeps feature-branch and fork uploads from
+        # requiring a repository secret.
+        if [ -n "${BUILDKITE_PULL_REQUEST:-}" ] \
+            && [ "${BUILDKITE_PULL_REQUEST}" != "false" ]; then
+            rust_cov_upload_branch="pr${BUILDKITE_PULL_REQUEST}:$rust_cov_upload_branch"
+        else
+            rust_cov_upload_branch="buildkite:$rust_cov_upload_branch"
+        fi
+    fi
+
+    set --
+    set -- "$@" upload-process
+    set -- "$@" --file "$rust_cov_upload_lcov"
+    # LCOV paths are mapped server-side by codecov.yml. Skip the CLI's local
+    # source-line fix scanning, which is unrelated to path mapping.
+    set -- "$@" --disable-search --disable-file-fixes
+    set -- "$@" --fail-on-error --git-service github
+    set -- "$@" --build "${BUILDKITE_BUILD_NUMBER:?BUILDKITE_BUILD_NUMBER is required}"
+    set -- "$@" --branch "$rust_cov_upload_branch"
+    set -- "$@" --sha "${BUILDKITE_COMMIT:?BUILDKITE_COMMIT is required}"
+    set -- "$@" --slug "$rust_cov_upload_slug"
+    set -- "$@" --flag "$rust_cov_upload_flag"
+    set -- "$@" --name "${rust_cov_upload_flag}-${BUILDKITE_JOB_ID:?BUILDKITE_JOB_ID is required}"
+    set -- "$@" --dir "$rust_cov_upload_repo_root"
+    set -- "$@" --network-root-folder "$rust_cov_upload_repo_root"
+    if [ -n "${BUILDKITE_PULL_REQUEST:-}" ] \
+        && [ "${BUILDKITE_PULL_REQUEST}" != "false" ]; then
+        set -- "$@" --pr "$BUILDKITE_PULL_REQUEST"
+    fi
+
+    rust_cov_upload_log="$rust_cov_upload_codecov_dir/codecov.log"
+    # E2E steps run from tests/, so execute from the repository root to resolve
+    # codecov.yml and repository paths consistently.
+    (
+        cd "$rust_cov_upload_repo_root" || exit 1
+        "$rust_cov_upload_codecov_dir/codecov" "$@"
+    ) >"$rust_cov_upload_log" 2>&1
+    rust_cov_upload_rc=$?
+    cat "$rust_cov_upload_log"
+    # v11.3.1 can log API failures while returning zero even with
+    # --fail-on-error. Preserve the strict CI contract explicitly.
+    if grep -aEq 'error.* -- ' "$rust_cov_upload_log"; then
+        echo "Codecov CLI reported an upload error" >&2
+        rust_cov_upload_rc=1
+    fi
+    rm -rf "$rust_cov_upload_codecov_dir"
+    return "$rust_cov_upload_rc"
+}
+
+rust_coverage_finalize() {
+    rust_cov_finalize_test_rc=$?
+    trap - 0
+    set +e
+
+    rust_coverage_collect "$RUST_COVERAGE_FLAG"
+    rust_cov_finalize_collect_rc=$?
+
+    rust_cov_finalize_upload_rc=0
+    if [ "$rust_cov_finalize_collect_rc" -eq 0 ]; then
+        rust_coverage_upload "$RUST_COVERAGE_LCOV" "$RUST_COVERAGE_FLAG"
+        rust_cov_finalize_upload_rc=$?
+    fi
+
+    find "$RUST_COVERAGE_DIR" -type f -name '*.profraw' -delete
+
+    if [ "$rust_cov_finalize_test_rc" -ne 0 ]; then
+        exit "$rust_cov_finalize_test_rc"
+    fi
+    if [ "$rust_cov_finalize_collect_rc" -ne 0 ]; then
+        exit "$rust_cov_finalize_collect_rc"
+    fi
+    exit "$rust_cov_finalize_upload_rc"
+}

.buildkite/test_areas/rust_frontend.yaml

@@ -8,6 +8,11 @@ steps:
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - rust/
+  - build_rust.sh
+  - tools/build_rust.py
+  - rust-toolchain.toml
+  - .buildkite/scripts/rust-coverage.sh
+  - codecov.yml
   - vllm/benchmarks/
   - vllm/entrypoints/openai/
   - vllm/entrypoints/serve/
@@ -23,6 +28,7 @@ steps:
   - tests/entrypoints/openai/test_uds.py
   - tests/v1/sample/test_logprobs_e2e.py
   commands:
+  - . /vllm-workspace/.buildkite/scripts/rust-coverage.sh && rust_coverage_start rust-e2e
   - export VLLM_USE_RUST_FRONTEND=1
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s benchmarks/test_serve_cli.py -k "not insecure and not (test_bench_serve and not test_bench_serve_chat)"
@@ -43,6 +49,11 @@ steps:
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - rust/
+  - build_rust.sh
+  - tools/build_rust.py
+  - rust-toolchain.toml
+  - .buildkite/scripts/rust-coverage.sh
+  - codecov.yml
   - vllm/entrypoints/openai/
   - vllm/entrypoints/serve/
   - vllm/v1/engine/
@@ -54,6 +65,7 @@ steps:
   # - tests/entrypoints/serve/dev/test_sleep.py
   - tests/entrypoints/serve/tokenize/test_tokenization.py
   commands:
+  - . /vllm-workspace/.buildkite/scripts/rust-coverage.sh && rust_coverage_start rust-e2e
   - export VLLM_USE_RUST_FRONTEND=1
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc/test_collective_rpc.py
@@ -72,10 +84,16 @@ steps:
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - rust/
+  - build_rust.sh
+  - tools/build_rust.py
+  - rust-toolchain.toml
+  - .buildkite/scripts/rust-coverage.sh
+  - codecov.yml
   - vllm/entrypoints/openai/
   - tests/utils.py
   - tests/entrypoints/openai/correctness/test_lmeval.py
   commands:
+  - . /vllm-workspace/.buildkite/scripts/rust-coverage.sh && rust_coverage_start rust-e2e
   - export VLLM_USE_RUST_FRONTEND=1
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
@@ -86,11 +104,17 @@ steps:
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - rust/
+  - build_rust.sh
+  - tools/build_rust.py
+  - rust-toolchain.toml
+  - .buildkite/scripts/rust-coverage.sh
+  - codecov.yml
   - vllm/entrypoints/openai/
   - vllm/tool_parsers/
   - tests/utils.py
   - tests/tool_use/
   commands:
+  - . /vllm-workspace/.buildkite/scripts/rust-coverage.sh && rust_coverage_start rust-e2e
   - export VLLM_USE_RUST_FRONTEND=1
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s tool_use --ignore=tool_use/mistral --models llama3.2 -k "not test_response_format_with_tool_choice_required and not test_parallel_tool_calls_false and not test_tool_call_and_choice"
@@ -101,6 +125,11 @@ steps:
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - rust/
+  - build_rust.sh
+  - tools/build_rust.py
+  - rust-toolchain.toml
+  - .buildkite/scripts/rust-coverage.sh
+  - codecov.yml
   - vllm/distributed/
   - vllm/engine/
   - vllm/executor/
@@ -111,6 +140,7 @@ steps:
   - tests/v1/distributed/test_hybrid_lb_dp.py
   - tests/v1/distributed/test_internal_lb_dp.py
   commands:
+  - . /vllm-workspace/.buildkite/scripts/rust-coverage.sh && rust_coverage_start rust-e2e
   - export VLLM_USE_RUST_FRONTEND=1
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - export NCCL_CUMEM_HOST_ENABLE=0

.buildkite/test_areas/rust_frontend_cargo.yaml

@@ -26,5 +26,7 @@ steps:
   - rust-toolchain.toml
   - .buildkite/test_areas/rust_frontend_cargo.yaml
   - .buildkite/scripts/run-rust-frontend-cargo-ci.sh
+  - .buildkite/scripts/rust-coverage.sh
+  - codecov.yml
   commands:
   - .buildkite/scripts/run-rust-frontend-cargo-ci.sh test

.gitignore

@@ -257,3 +257,4 @@ vllm/grpc/vllm_engine_pb2.pyi
 
 # Ignore generated cpu headers 
 csrc/cpu/cpu_attn_dispatch_generated.h
+rust-coverage-tools/

build_rust.sh

@@ -8,6 +8,8 @@
 set -euo pipefail
 
 REPO_ROOT="$(cd "$(dirname "$0")" && pwd)"
+CARGO_LLVM_COV_VERSION="0.8.7"
+COVERAGE_TOOLS_DIR="$REPO_ROOT/rust-coverage-tools"
 
 # Read the required toolchain from rust-toolchain.toml.
 TOOLCHAIN=$(grep '^channel' "$REPO_ROOT/rust-toolchain.toml" | sed 's/.*= *"\(.*\)"/\1/')
@@ -30,4 +32,39 @@ else
     PROFILE_ARG="--release"
 fi
 
+rm -rf "$COVERAGE_TOOLS_DIR"
+mkdir -p "$COVERAGE_TOOLS_DIR/bin" "$COVERAGE_TOOLS_DIR/lib"
+
+if [[ "${VLLM_RUST_COVERAGE:-0}" == "1" ]]; then
+    # rustc wrapper flags are invisible to Cargo's normal fingerprinting.
+    # Keep instrumented intermediates isolated when local builds switch modes.
+    export CARGO_TARGET_DIR="$REPO_ROOT/rust/target/coverage"
+    rustup component add --toolchain "$TOOLCHAIN" llvm-tools-preview
+    cargo +"$TOOLCHAIN" install \
+        --locked \
+        --version "$CARGO_LLVM_COV_VERSION" \
+        cargo-llvm-cov
+
+    eval "$(
+        cargo +"$TOOLCHAIN" llvm-cov show-env \
+            --manifest-path "$REPO_ROOT/rust/Cargo.toml" \
+            --sh
+    )"
+
+    # Build scripts and proc macros can run during compilation. Their profiles
+    # are unrelated to runtime coverage and would otherwise pollute the tree.
+    export LLVM_PROFILE_FILE=/dev/null
+    export VLLM_RUST_COVERAGE_OBJECTS="$COVERAGE_TOOLS_DIR/objects"
+fi
+
 python3 "$REPO_ROOT/tools/build_rust.py" "$PROFILE_ARG"
+
+if [[ "${VLLM_RUST_COVERAGE:-0}" == "1" ]]; then
+    LLVM_BIN_DIR="$(dirname "$(rustup run "$TOOLCHAIN" rustc \
+        --print target-libdir)")/bin"
+
+    cp "$LLVM_BIN_DIR"/{llvm-cov,llvm-profdata} "$COVERAGE_TOOLS_DIR/bin/"
+    chmod 0755 "$COVERAGE_TOOLS_DIR/bin/"*
+    cp -L "$LLVM_BIN_DIR"/../lib/libLLVM.so* "$COVERAGE_TOOLS_DIR/lib/"
+    chmod 0644 "$COVERAGE_TOOLS_DIR/lib/"*
+fi

codecov.yml

@@ -10,3 +10,16 @@ fixes:
   - "/usr/local/lib/python3.*/site-packages/vllm/::vllm/"
   - "/usr/lib/python3.*/dist-packages/vllm/::vllm/"
   - "/usr/lib/python3.*/site-packages/vllm/::vllm/"
+  # Map Rust sources built in the E2E image and on Buildkite agents.
+  - "/workspace/rust/::rust/"
+  - "/var/lib/buildkite-agent/.*/rust/::rust/"
+
+flags:
+  rust-unit:
+    paths:
+      - rust/
+    carryforward: false
+  rust-e2e:
+    paths:
+      - rust/
+    carryforward: false

csrc/cpu/cpu_attn_vec.hpp

@@ -102,7 +102,9 @@ class TileGemm82 {
     kv_cache_t* __restrict__ curr_b = b_tile;
 
     for (int32_t k = 0; k < dynamic_k_size; ++k) {
-      auto [fp32_b_0_reg, fp32_b_1_reg] = load_b_pair_vec(curr_b);
+      auto fp32_b_regs = load_b_pair_vec(curr_b);
+      auto fp32_b_0_reg = fp32_b_regs.first;
+      auto fp32_b_1_reg = fp32_b_regs.second;
 
       float* __restrict__ curr_m_a = curr_a;
       vec_op::unroll_loop<int32_t, M>([&](int32_t i) {

docker/Dockerfile

@@ -294,6 +294,9 @@ FROM base AS rust-build
 ARG BUILD_OS
 ARG USE_SCCACHE
 ARG SCCACHE_ENDPOINT
+# Temporary default for the initial CI validation. Set this back to 0 when
+# ci-infra passes VLLM_RUST_COVERAGE=1 explicitly.
+ARG VLLM_RUST_COVERAGE=1
 
 # Install native tools needed only for Rust/protoc builds.
 RUN if [ "${BUILD_OS}" = "manylinux" ]; then \
@@ -902,6 +905,14 @@ COPY ./vllm/collect_env.py .
 # note that this uses vllm installed by `pip`
 FROM vllm-base AS test
 
+COPY --from=rust-build \
+    /workspace/rust-coverage-tools/ \
+    /opt/vllm-rust-coverage/
+
+ENV PATH=/opt/vllm-rust-coverage/bin:${PATH}
+ENV LD_LIBRARY_PATH=/opt/vllm-rust-coverage/lib:${LD_LIBRARY_PATH}
+ENV LLVM_PROFILE_FILE=/dev/null
+
 ADD . /vllm-workspace/
 
 ARG PYTHON_VERSION

docker/versions.json

@@ -46,6 +46,9 @@
     "TORCH_CUDA_ARCH_LIST": {
       "default": "7.5 8.0 8.6 8.9 9.0 10.0 11.0 12.0"
     },
+    "VLLM_RUST_COVERAGE": {
+      "default": "1"
+    },
     "MAX_JOBS": {
       "default": "2"
     },

tools/build_rust.py

@@ -36,6 +36,22 @@ def rust_extensions(*, optional: bool = False) -> list[RustExtension]:
     ]
 
 
+def write_coverage_objects(extensions: list[RustExtension], output: Path) -> None:
+    artifacts = []
+    for extension in extensions:
+        for target in sorted(set(extension.target.values())):
+            target_path = ROOT_DIR.joinpath(*target.split("."))
+            if extension.binding == Binding.Exec:
+                matches = [target_path]
+            else:
+                matches = sorted(target_path.parent.glob(f"{target_path.name}*.so"))
+            if len(matches) != 1 or not matches[0].is_file():
+                raise RuntimeError(f"unable to locate Rust artifact for {target}")
+            artifacts.append(matches[0].relative_to(ROOT_DIR).as_posix())
+
+    output.write_text("\n".join(artifacts) + "\n")
+
+
 def rust_py_extension_module_names() -> list[str]:
     module_names = []
     for extension in rust_extensions():
@@ -52,12 +68,15 @@ def rust_py_extension_module_names() -> list[str]:
 def build_binary(build_rust_args: list[str]) -> None:
     os.chdir(ROOT_DIR)
     (ROOT_DIR / "vllm").mkdir(exist_ok=True)
+    extensions = rust_extensions(optional=False)
     setup(
         name="vllm-rust-frontend-build",
         packages=[],
-        rust_extensions=rust_extensions(optional=False),
+        rust_extensions=extensions,
         script_args=["build_rust", "--quiet", "--inplace", *build_rust_args],
     )
+    if output := os.getenv("VLLM_RUST_COVERAGE_OBJECTS"):
+        write_coverage_objects(extensions, Path(output))
 
 
 def main() -> None:

vllm/entrypoints/openai/chat_completion/protocol.py

@@ -687,9 +687,9 @@ class ChatCompletionRequest(OpenAIBaseModel):
             skip_special_tokens=self.skip_special_tokens,
             spaces_between_special_tokens=self.spaces_between_special_tokens,
             include_stop_str_in_output=self.include_stop_str_in_output,
-            output_kind=RequestOutputKind.DELTA
-            if self.stream
-            else RequestOutputKind.FINAL_ONLY,
+            output_kind=(
+                RequestOutputKind.DELTA if self.stream else RequestOutputKind.FINAL_ONLY
+            ),
             structured_outputs=self.extract_structured_outputs(),
             logit_bias=self.logit_bias,
             bad_words=self.bad_words,
@@ -849,9 +849,10 @@ class ChatCompletionRequest(OpenAIBaseModel):
 
         # Reject empty tools array, matching OpenAI API behavior
         if data.get("tools") == []:
-            raise ValueError(
+            raise VLLMValidationError(
                 "`tools` must not be an empty array. "
-                "Either provide at least one tool or omit the field entirely."
+                "Either provide at least one tool or omit the field entirely.",
+                parameter="tools",
             )
 
         # if "tool_choice" is not specified but tools are provided,
@@ -1075,13 +1076,15 @@ class BatchChatCompletionRequest(OpenAIBaseModel):
         if isinstance(data, BatchChatCompletionRequest):
             data = data.model_dump(exclude_unset=True)
         if data.get("use_beam_search"):
-            raise ValueError(
+            raise VLLMValidationError(
                 "Batch chat completions do not support beam search. "
-                "Please set `use_beam_search` to False."
+                "Please set `use_beam_search` to False.",
+                parameter="use_beam_search",
             )
         if data.get("logprob_token_ids") and not data.get("logprobs"):
-            raise ValueError(
-                "when using `logprob_token_ids`, `logprobs` must be set to true."
+            raise VLLMValidationError(
+                "when using `logprob_token_ids`, `logprobs` must be set to true.",
+                parameter="logprob_token_ids",
             )
         response_format = data.get("response_format")
         rf_type = (
@@ -1095,8 +1098,10 @@ class BatchChatCompletionRequest(OpenAIBaseModel):
             validate_structured_outputs_structural_tag(structured_outputs)
         n = data.get("n", 1)
         if n is not None and n != 1:
-            raise ValueError(
-                "Batch chat completions do not support `n > 1`. Please set `n` to 1."
+            raise VLLMValidationError(
+                "Batch chat completions do not support `n > 1`. Please set `n` to 1.",
+                parameter="n",
+                value=n,
             )
         return data