[CI] Discover Rust coverage artifacts from build metadata

Co-authored-by: OpenAI Codex <codex@openai.com>
[CI] Simplify Rust coverage reporting
2026-07-23 06:44:16 +00:00 · 2026-07-22 08:18:14 +00:00 · 2026-07-22 02:58:45 +00:00 · 2026-07-21 13:57:38 +00:00 · 2026-07-21 13:34:22 +00:00 · 2026-07-21 13:01:56 +00:00
202 changed files with 6527 additions and 4444 deletions
@@ -813,8 +813,8 @@ steps:

            # Download artifacts from current build
            echo "Downloading artifacts from current build"
-            # buildkite-agent artifact download "artifacts/rocm-base-wheels/*.whl" .
-            # buildkite-agent artifact download "artifacts/rocm-vllm-wheel/*.whl" .
+            buildkite-agent artifact download "artifacts/rocm-base-wheels/*.whl" .
+            buildkite-agent artifact download "artifacts/rocm-vllm-wheel/*.whl" .

            # # Run upload script
            bash .buildkite/scripts/upload-rocm-wheels.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
@@ -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"
+}
@@ -10,7 +10,9 @@ steps:
  - vllm/engine/arg_utils.py
  - vllm/config/model.py
  - vllm/model_executor
+  - vllm/model_executor/warmup
  - tests/model_executor
+  - tests/model_executor/test_jit_warmup.py
  - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
  commands:
    - apt-get update && apt-get install -y curl libsodium23
@@ -34,7 +36,9 @@ steps:
      - vllm/engine/arg_utils.py
      - vllm/config/model.py
      - vllm/model_executor
+      - vllm/model_executor/warmup
      - tests/model_executor
+      - tests/model_executor/test_jit_warmup.py
      - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
      - vllm/_aiter_ops.py
      - vllm/platforms/rocm.py
@@ -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
@@ -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
@@ -47,6 +47,7 @@

 # Rust Frontend
 /rust/ @BugenZhao @njhill
+/rust/src/bench @esmeetu
 /build_rust.sh @BugenZhao @njhill
 /rust-toolchain.toml @BugenZhao @njhill
 /.buildkite/test_areas/rust* @BugenZhao @njhill
@@ -257,3 +257,4 @@ vllm/grpc/vllm_engine_pb2.pyi

 # Ignore generated cpu headers 
 csrc/cpu/cpu_attn_dispatch_generated.h
+rust-coverage-tools/
@@ -48,7 +48,7 @@ vLLM is flexible and easy to use with:
 - Tool calling and reasoning parsers
 - OpenAI-compatible API server, plus Anthropic Messages API and gRPC support
 - Efficient multi-LoRA support for dense and MoE layers
- Support for NVIDIA GPUs, AMD GPUs, and x86/ARM/PowerPC CPUs. Additionally, diverse hardware plugins such as Google TPUs, Intel Gaudi, IBM Spyre, Huawei Ascend, Rebellions NPU, Apple Silicon, MetaX GPU, and more.
+- Support for NVIDIA GPUs, AMD GPUs, Intel GPUs, and x86/ARM/PowerPC CPUs. Additionally, diverse hardware plugins such as Google TPUs, Intel Gaudi, IBM Spyre, Huawei Ascend, Rebellions NPU, Apple Silicon, MetaX GPU, and more.

 vLLM seamlessly supports 200+ model architectures on Hugging Face, including:

@@ -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
@@ -22,7 +22,7 @@ if(QUTLASS_SRC_DIR)
  set(qutlass_BINARY_DIR "${CMAKE_BINARY_DIR}/qutlass-binary-dir-unused")
 else()
  set(_QUTLASS_UPSTREAM_REPO "https://github.com/IST-DASLab/qutlass.git")
-  set(_QUTLASS_UPSTREAM_TAG "830d2c4537c7396e14a02a46fbddd18b5d107c65")
+  set(_QUTLASS_UPSTREAM_TAG "e74319e3405ce6d71965732880f5dc1f52371f64")

  set(_qutlass_fc_root "${FETCHCONTENT_BASE_DIR}")
  if(NOT _qutlass_fc_root)
@@ -125,8 +125,6 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
    CUDA_ARCHS "${QUTLASS_ARCHS}"
  )

-  # QuTLASS uses legacy ATen headers and cannot be built with TORCH_TARGET_VERSION.
-  # Keep it as its own extension (registers torch.ops._qutlass_C).
  define_extension_target(
    _qutlass_C
    DESTINATION vllm
@@ -139,9 +137,11 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
    WITH_SOABI)

  target_compile_definitions(_qutlass_C PRIVATE
-    QUTLASS_DISABLE_PYBIND=1
+    QUTLASS_MINIMAL_BUILD=1
    TARGET_CUDA_ARCH=${QUTLASS_TARGET_CC}
-    CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
+    CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1
+    TORCH_TARGET_VERSION=0x020B000000000000ULL
+    USE_CUDA)

  set_property(SOURCE ${QUTLASS_SOURCES} APPEND PROPERTY COMPILE_OPTIONS
    $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr --use_fast_math -O3>
@@ -39,7 +39,7 @@ else()
  FetchContent_Declare(
          vllm-flash-attn
          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG caaa4eb59845388a20b1f435ecaafb4bd9517ad8
+          GIT_TAG 168920233059c48de6199e2cda74003b2ce3d199
          GIT_PROGRESS TRUE
          # Don't share the vllm-flash-attn build between build types
          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
@@ -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
@@ -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) {
@@ -39,11 +39,15 @@ __global__ void marlin_int4_fp8_preprocess_kernel_awq(
    // AWQ zeros: (size_k // group_size, size_n // 8)
    const int32_t* __restrict__ qzeros, int32_t size_n, int32_t size_k,
    int32_t group_size) {
-  int32_t val =
-      qweight[(blockIdx.x * 32 + threadIdx.x) * size_n / 8 + blockIdx.y];
-  int32_t zero =
-      qzeros[(blockIdx.x * 32 + threadIdx.x) / group_size * size_n / 8 +
-             blockIdx.y];
+  // Thread mapping: threadIdx.x -> column dim (coalesced read within a row),
+  // blockIdx.x -> row dim. Adjacent threads read consecutive int32 in the
+  // same row (stride 1) instead of striding across rows (stride size_n/8).
+  int col = blockIdx.y * 32 + threadIdx.x;
+  if (col >= size_n / 8) return;
+  (void)size_k;
+
+  int32_t val = qweight[blockIdx.x * (size_n / 8) + col];
+  int32_t zero = qzeros[blockIdx.x / group_size * (size_n / 8) + col];
  int32_t new_val = 0;

 #pragma unroll
@@ -58,7 +62,7 @@ __global__ void marlin_int4_fp8_preprocess_kernel_awq(
    zero >>= 4;
  }

-  output[(blockIdx.x * 32 + threadIdx.x) * size_n / 8 + blockIdx.y] = new_val;
+  output[blockIdx.x * (size_n / 8) + col] = new_val;
 }

 torch::stable::Tensor marlin_int4_fp8_preprocess(
@@ -102,7 +106,7 @@ torch::stable::Tensor marlin_int4_fp8_preprocess(
                    "qweight.size(0) % qzeros.size(0) != 0");
    STD_TORCH_CHECK(group_size % 8 == 0, "group_size % 8 != 0");

-    dim3 blocks(size_k / 32, size_n / 8);
+    dim3 blocks(size_k, (size_n / 8 + 31) / 32);
    marlin_int4_fp8_preprocess_kernel_awq<<<blocks, 32, 0, stream>>>(
        reinterpret_cast<const int32_t*>(qweight.const_data_ptr()),
        reinterpret_cast<int32_t*>(output.mutable_data_ptr()),
@@ -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
@@ -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"
    },
@@ -164,8 +164,8 @@ Priority is **1 = highest** (tried first).
 | `FLASHINFER` | XQA† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ❌ | ❌ | ❌ | ✅ | Decoder | 9.0 |
 | `FLASHINFER` | trtllm-gen† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `nvfp4` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ✅ | ✅ | ❌ | ✅ | Decoder | 10.x |
 | `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ❌ | ✅ | All | ≥8.0 |
-| `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | 9.x |
-| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | ≥10.0 |
+| `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | 9.x |
+| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | ≥10.0 |
 | `FLASH_ATTN_DIFFKV` | | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
 | `FLEX_ATTENTION` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ✅ | ❌ | Decoder, Encoder Only | Any |
 | `HPC_ATTN` | | fp16, bf16 | `auto`, `bfloat16`, `fp8_e4m3` | 64 | 128 | ❌ | ❌ | ❌ | ❌ | Decoder | ≥9.0 |
@@ -75,7 +75,7 @@ vllm serve <model> \
 | `max_tracker_size` | no | `64000` | single-tier | Max entries in the lookup tracker. |
 | `secondary_tiers` | no | `[]` | multi-tier | List of secondary tier configs (see below). |
 | `offload_prompt_only` | no | `true` | both | If `true`, only prompt (prefill) blocks are offloaded; decode blocks are skipped. |
-| `self_describing_kv_events` | no | `false` | single-tier | Opt-in. When `true` *and* KV cache events are enabled (`--kv-events-config` with `enable_kv_cache_events`), the connector emits self-describing block-granular `BlockStored`/`BlockRemoved` payloads (constituent block hashes, whole-chunk `token_ids`, per-block `block_size`, parent hash, LoRA + group/cache-spec metadata) instead of the placeholder fallback, so external KV-event consumers can index offloaded blocks. Inert unless events are enabled. Currently rejected by `TieringOffloadingSpec`. Full-attention groups only; sliding-window/SSM groups keep the placeholder fallback. In chunk mode (`block_size` > GPU block size, or `blocks_per_chunk` > 1), overlapping chunks re-announce shared per-block hashes, so consumers must reference-count (deduplicate) repeated store/remove announcements. |
+| `self_describing_kv_events` | no | `false` | both | Opt-in. When `true` *and* KV cache events are enabled (`--kv-events-config` with `enable_kv_cache_events`), the connector emits self-describing block-granular `BlockStored`/`BlockRemoved` payloads (constituent block hashes, whole-chunk `token_ids`, per-block `block_size`, parent hash, LoRA + group/cache-spec metadata) instead of the placeholder fallback, so external KV-event consumers can index offloaded blocks. Inert unless events are enabled. With `TieringOffloadingSpec`, a CPU promotion is self-describing when a local request observes its primary-tier `HIT` before event translation; otherwise its stored event may retain the placeholder, while a later `HIT` can backfill metadata for removal. Pending-removal/re-promotion races and externally initiated promotions may also produce placeholders, and consumers must ignore removals for unknown hashes. Full-attention groups only; sliding-window/SSM groups keep the placeholder fallback. In chunk mode (`block_size` > GPU block size, or `blocks_per_chunk` > 1), overlapping chunks re-announce shared per-block hashes, so consumers must reference-count (deduplicate) repeated store/remove announcements. |
 | `spec_module_path` | no | — | both | Python import path for a custom `OffloadingSpec` not in the built-in registry. Required only when `spec_name` is not built-in (advanced). |

 ## Secondary Tiers
@@ -27,7 +27,7 @@ Currently, there are no pre-built XPU wheels.

 - First, install required [driver](https://dgpu-docs.intel.com/driver/installation.html#installing-gpu-drivers).
 - Second, install Python packages for vLLM XPU backend building (Intel OneAPI dependencies are installed automatically as part of `torch-xpu`, see [PyTorch XPU get started](https://docs.pytorch.org/docs/stable/notes/get_start_xpu.html)):
- Start from vllm-xpu-kernels v0.1.10, we recommend user upgrade driver to [compute runtime 26.18](https://github.com/intel/compute-runtime/releases/tag/26.14.37833.4) release, to avoid potential compatibility issue.
+- Start from vllm-xpu-kernels v0.1.10, we recommend user upgrade driver to [compute runtime 26.18](https://github.com/intel/compute-runtime/releases/tag/26.18.38308.1) release, to avoid potential compatibility issue.

 ```bash
 git clone https://github.com/vllm-project/vllm.git
@@ -58,7 +58,40 @@ VLLM_TARGET_DEVICE=xpu pip install --no-build-isolation -e . -v
 --8<-- [end:build-wheel-from-source]
 --8<-- [start:pre-built-images]

-Currently, we release prebuilt XPU images at docker [hub](https://hub.docker.com/r/intel/vllm/tags) based on vLLM released version. For more information, please refer release [note](https://github.com/intel/ai-containers/blob/main/vllm).
+vLLM offers official Docker images for deployment.
+The images can be used to run OpenAI compatible server and are available on Docker Hub as [vllm/vllm-openai-xpu](https://hub.docker.com/r/vllm/vllm-openai-xpu/tags).
+
+- `vllm/vllm-openai-xpu:latest` — stable release, available starting from v0.26.0
+- `vllm/vllm-openai-xpu:nightly` — preview build from the latest development branch, use this if you want the latest features and fixes
+
+```bash
+docker run --rm \
+    --network=host \
+    --device /dev/dri:/dev/dri \
+    -v /dev/dri/by-path:/dev/dri/by-path \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    --env "HF_TOKEN=$HF_TOKEN" \
+    --ipc=host \
+    --privileged \
+    vllm/vllm-openai-xpu:<tag> \
+    --model Qwen/Qwen3-0.6B
+```
+
+To use the docker image as base for development, you can launch it in interactive session through overriding the entrypoint.
+
+???+ console "Commands"
+    ```bash
+    docker run --rm -it \
+        --network=host \
+        --device /dev/dri:/dev/dri \
+        -v /dev/dri/by-path:/dev/dri/by-path \
+        -v ~/.cache/huggingface:/root/.cache/huggingface \
+        --env "HF_TOKEN=$HF_TOKEN" \
+        --ipc=host \
+        --privileged \
+        --entrypoint /bin/bash \
+        vllm/vllm-openai-xpu:<tag>
+    ```

 --8<-- [end:pre-built-images]
 --8<-- [start:build-image-from-source]
@@ -65,6 +65,15 @@ This guide will help you quickly get started with vLLM to perform:
    !!! tip
        A nightly Docker image is also available as [vllm/vllm-openai-rocm:nightly](https://hub.docker.com/r/vllm/vllm-openai-rocm/tags) for testing the latest development builds.

+=== "Intel GPU"
+
+    vLLM supports Intel GPUs through the XPU backend. Pre-built XPU wheels will be available soon.
+
+    Official Docker images for Intel GPUs are added to the vLLM release starting from v0.26.0. Nightly Docker image is also available as [vllm/vllm-openai-xpu:nightly](https://hub.docker.com/r/vllm/vllm-openai-xpu/tags).
+
+    !!! tip
+        For more detailed instructions, including building from source and Docker image setup, please refer to the [GPU installation guide](installation/gpu.md) and select the "Intel XPU" tab.
+
 === "Google TPU"

    To run vLLM on Google TPUs, you need to install the `vllm-tpu` package.
@@ -67,7 +67,7 @@ The Transcriptions API supports uploading audio files in various formats includi
 - `response_format`: Format of the response ("json", "text") (optional)
 - `temperature`: Sampling temperature between 0 and 1 (optional)

-For the complete list of supported parameters including sampling parameters and vLLM extensions, see the [protocol definitions](https://github.com/vllm-project/vllm/blob/main/vllm/entrypoints/openai/protocol.py#L2182).
+For the complete list of supported parameters including sampling parameters and vLLM extensions, see the [protocol definitions](https://github.com/vllm-project/vllm/blob/main/vllm/entrypoints/speech_to_text/transcription/protocol.py).

 **Response Format:**

@@ -3446,7 +3446,6 @@ dependencies = [
 "base64 0.22.1",
 "bytes",
 "encoding_rs",
- "futures-channel",
 "futures-core",
 "futures-util",
 "h2",
@@ -4876,6 +4875,7 @@ dependencies = [
 "futures-core",
 "pin-project-lite",
 "tokio",
+ "tokio-util",
 ]

 [[package]]
@@ -4937,9 +4937,9 @@ dependencies = [

 [[package]]
 name = "tonic"
-version = "0.14.5"
+version = "0.14.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fec7c61a0695dc1887c1b53952990f3ad2e3a31453e1f49f10e75424943a93ec"
+checksum = "ac2a5518c70fa84342385732db33fb3f44bc4cc748936eb5833d2df34d6445ef"
 dependencies = [
 "async-trait",
 "axum",
@@ -4966,9 +4966,9 @@ dependencies = [

 [[package]]
 name = "tonic-build"
-version = "0.14.5"
+version = "0.14.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1882ac3bf5ef12877d7ed57aad87e75154c11931c2ba7e6cde5e22d63522c734"
+checksum = "c68f61875ac5293cf72e6c8cf0158086428c82c37229e98c840878f1706b0322"
 dependencies = [
 "prettyplease",
 "proc-macro2",
@@ -4977,10 +4977,23 @@ dependencies = [
 ]

 [[package]]
-name = "tonic-prost"
-version = "0.14.5"
+name = "tonic-health"
+version = "0.14.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a55376a0bbaa4975a3f10d009ad763d8f4108f067c7c2e74f3001fb49778d309"
+checksum = "fcfab99db777fba2802f0dfa861d1628d1ae916fb199d29819941f139ae85082"
+dependencies = [
+ "prost",
+ "tokio",
+ "tokio-stream",
+ "tonic",
+ "tonic-prost",
+]
+
+[[package]]
+name = "tonic-prost"
+version = "0.14.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "50849f68853be452acf590cde0b146665b8d507b3b8af17261df47e02c209ea0"
 dependencies = [
 "bytes",
 "prost",
@@ -4989,9 +5002,9 @@ dependencies = [

 [[package]]
 name = "tonic-prost-build"
-version = "0.14.5"
+version = "0.14.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f3144df636917574672e93d0f56d7edec49f90305749c668df5101751bb8f95a"
+checksum = "654e5643eff75d7f8c99197ce1440ed19a3474eada74c12bbac488b2cafdae27"
 dependencies = [
 "prettyplease",
 "proc-macro2",
@@ -5484,10 +5497,13 @@ dependencies = [
 "serde",
 "serde_json",
 "thiserror 2.0.18",
+ "thiserror-ext",
 "tiktoken-rs 0.9.1",
 "tokenizers",
 "tokio",
 "tokio-stream",
+ "tracing",
+ "tracing-subscriber",
 "url",
 "uuid",
 ]
@@ -5730,6 +5746,7 @@ dependencies = [
 "tokio-stream",
 "tokio-util",
 "tonic",
+ "tonic-health",
 "tonic-prost",
 "tonic-prost-build",
 "tower",
@@ -6321,9 +6338,9 @@ checksum = "9edde0db4769d2dc68579893f2306b26c6ecfbe0ef499b013d731b7b9247e0b9"

 [[package]]
 name = "xgrammar-structural-tag"
-version = "0.1.0+xgrammar.0.2.2.4d145cc"
+version = "0.2.0+xgrammar.0.2.4.dd729e7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2436dea2393d55a3b188588aa300c5a8afe8f45a77da52c611fb4498a6c876e6"
+checksum = "d4d24c842efc3c24e9756aa426d530cbdac0980e49af223cb384e276e981ca0a"
 dependencies = [
 "auto_impl",
 "serde",
@@ -118,10 +118,11 @@ tokio = { version = "1.47.1", features = [
 tokio-openssl = "0.6"
 tokio-stream = "0.1"
 tokio-util = { version = "0.7.18", features = ["rt"] }
-tonic = "0.14.5"
-tonic-build = "0.14.5"
-tonic-prost = "0.14.5"
-tonic-prost-build = "0.14.5"
+tonic = "0.14.6"
+tonic-build = "0.14.6"
+tonic-health = "0.14.6"
+tonic-prost = "0.14.6"
+tonic-prost-build = "0.14.6"
 tool-parser = "1.2.0"
 tower = { version = "0.5.3", features = ["util"] }
 tower-http = { version = "0.6.8", features = ["cors", "trace"] }
@@ -143,7 +144,7 @@ vllm-server = { path = "src/server" }
 vllm-text = { path = "src/text" }
 vllm-tokenizer = { path = "src/tokenizer" }
 winnow = { version = "1.0.2", features = ["simd"] }
-xgrammar-structural-tag = "0.1.0"
+xgrammar-structural-tag = "0.2.0"
 zeromq = { version = "0.6.0", default-features = false, features = [
    "tokio-runtime",
    "all-transport",
@@ -20,16 +20,19 @@ mimalloc.workspace = true
 rand.workspace = true
 rand_distr.workspace = true
 rayon.workspace = true
-reqwest = { workspace = true, features = ["json", "stream", "blocking", "http2"] }
+reqwest = { workspace = true, features = ["json", "stream", "http2"] }
 rlimit.workspace = true
 rustc-hash.workspace = true
 serde = { workspace = true, features = ["rc"] }
 serde_json = { workspace = true, features = ["raw_value"] }
 thiserror.workspace = true
+thiserror-ext.workspace = true
 tiktoken-rs.workspace = true
 tokenizers.workspace = true
 tokio.workspace = true
 tokio-stream.workspace = true
+tracing.workspace = true
+tracing-subscriber.workspace = true
 url.workspace = true
 uuid.workspace = true

@@ -158,9 +158,10 @@ impl PoolingBackend {
                // (mirrors Python async_request_vllm_rerank).
                if let Some(ref list) = input.prompt_list {
                    if list.len() < 2 {
-                        eprintln!(
-                            "WARNING: vllm-rerank request has no documents \
-                             (prompt_list needs [query, doc, ...])"
+                        tracing::warn!(
+                            backend = "vllm-rerank",
+                            inputs = list.len(),
+                            "rerank request has no documents"
                        );
                    }
                    let query = list.first().map(|s| s.as_ref()).unwrap_or("");
@@ -175,10 +176,10 @@ impl PoolingBackend {
                    // Legacy path: text prompt as query, documents via --extra-body.
                    let query = input.prompt.as_ref();
                    if query.is_empty() && input.prompt_token_ids.is_some() {
-                        eprintln!(
-                            "WARNING: vllm-rerank received empty query (random dataset uses \
-                             token IDs only). Use --dataset-name random-rerank for meaningful \
-                             rerank benchmarks."
+                        tracing::warn!(
+                            backend = "vllm-rerank",
+                            dataset = "random",
+                            "rerank request has an empty query; use the random-rerank dataset"
                        );
                    }
                    serde_json::json!({
@@ -6,6 +6,7 @@ use std::sync::Arc;
 use std::time::Instant;

 use indicatif::{ProgressBar, ProgressStyle};
+use thiserror_ext::AsReport as _;
 use tokio::sync::Semaphore;

 use crate::backends::{RequestFuncInput, RequestFuncOutput, get_backend};
@@ -72,12 +73,12 @@ pub fn pre_resolve_dns(
            v4.extend(v6);
            if !v4.is_empty() {
                let ips: Vec<_> = v4.iter().map(|a| a.ip()).collect();
-                println!("Pre-resolved {host} -> {ips:?}");
+                tracing::info!(host, addresses = ?ips, "pre-resolved benchmark endpoint DNS");
                builder = builder.resolve_to_addrs(host, &v4);
            }
        }
        Err(e) => {
-            eprintln!("Warning: DNS pre-resolution for '{host}' failed: {e}");
+            tracing::warn!(host, error = %e.as_report(), "failed to pre-resolve benchmark endpoint DNS");
        }
    }

@@ -346,10 +347,14 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
    let (model_id, model_name) = if let Some(ref m) = config.model {
        (m.clone(), config.model_name.clone())
    } else {
-        println!("Model not specified, fetching first model from server...");
+        tracing::info!(base_url = %config.base_url, "fetching first model from server");
        let (name, id) =
            get_first_model_from_server(&config.base_url, &client, &config.extra_headers).await?;
-        println!("First model name: {name}, first model id: {id}");
+        tracing::info!(
+            model_name = name,
+            model_id = id,
+            "selected first model from server"
+        );
        (id, Some(name))
    };

@@ -358,10 +363,10 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
        None
    } else {
        let tid = config.tokenizer_id.as_deref().unwrap_or(&model_id);
-        println!("Loading tokenizer: {tid}");
+        tracing::info!(tokenizer = tid, "loading tokenizer");
        let server_info = Some((config.base_url.as_str(), model_id.as_str()));
-        let t = crate::tokenizer::load_tokenizer(tid, config.trust_remote_code, server_info)?;
-        println!("Tokenizer loaded successfully.");
+        let t =
+            crate::tokenizer::load_tokenizer(tid, config.trust_remote_code, server_info).await?;
        Some(t)
    };
    let has_tokenizer = tokenizer.is_some();
@@ -421,7 +426,12 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
            config.num_prompts, config.random_batch_size, config.is_reranker,
        ),
    };
-    println!("Generating {dataset_label}...");
+    tracing::info!(
+        dataset = ?config.dataset_name,
+        prompts = config.num_prompts,
+        description = %dataset_label,
+        "generating benchmark dataset"
+    );
    let gen_start = Instant::now();

    let mut input_requests = match config.dataset_name {
@@ -472,7 +482,7 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
            let path = match config.dataset_path.as_deref() {
                Some(p) => p,
                None => {
-                    downloaded = crate::datasets::sharegpt::download_sharegpt_dataset()?;
+                    downloaded = crate::datasets::sharegpt::download_sharegpt_dataset().await?;
                    downloaded.as_str()
                }
            };
@@ -512,7 +522,8 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
                None => {
                    downloaded = crate::datasets::speed_bench::download_speed_bench(
                        config.speed_bench_config,
-                    )?;
+                    )
+                    .await?;
                    downloaded.as_str()
                }
            };
@@ -543,7 +554,8 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
                    config.hf_subset.as_deref(),
                    config.hf_split.as_deref(),
                    config.num_prompts,
-                )?;
+                )
+                .await?;
            crate::datasets::hf_dataset::load_hf_dataset(
                tok,
                &downloaded_path,
@@ -608,18 +620,19 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
    };

    let gen_elapsed = gen_start.elapsed();
-    println!(
-        "Generated {} prompts in {:.2}s",
-        input_requests.len(),
-        gen_elapsed.as_secs_f64()
+    tracing::info!(
+        prompts = input_requests.len(),
+        elapsed_seconds = gen_elapsed.as_secs_f64(),
+        "generated benchmark dataset"
    );

    let filtered_count =
        filter_requests_by_max_model_len(&mut input_requests, config.max_model_len);
    if filtered_count > 0 {
-        println!(
-            "Filtered {filtered_count} prompt(s) above --max-model-len {}.",
-            config.max_model_len.unwrap()
+        tracing::info!(
+            filtered_prompts = filtered_count,
+            max_model_len = config.max_model_len.unwrap(),
+            "filtered prompts above maximum model length"
        );
    }
    if input_requests.is_empty() {
@@ -670,7 +683,7 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {

    // Ready check
    if config.ready_check_timeout_sec > 0 {
-        println!("Starting initial single prompt test run...");
+        tracing::info!("starting initial single-prompt test run");
        let test_output = wait_for_endpoint(
            config.backend,
            &client,
@@ -685,7 +698,7 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
                test_output.error
            )));
        }
-        println!("Initial test run completed.");
+        tracing::info!("initial single-prompt test run completed");
    }

    // Verify and fix prompt token lengths against the server's /tokenize endpoint.
@@ -703,12 +716,15 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
        DatasetName::Random | DatasetName::PrefixRepetition
    );
    if verifiable_dataset && has_token_ids && !config.backend.is_pooling() {
-        println!("Using prompt_token_ids, skipping server-side tokenizer verification.");
+        tracing::info!(
+            reason = "prompt_token_ids",
+            "skipping server tokenizer verification"
+        );
    }
    if verifiable_dataset && !has_token_ids && !config.backend.is_pooling() {
        let cache_key = tokenizer_verify_cache_key(&config.base_url, &model_id);
        if is_tokenizer_verified(&cache_key) {
-            println!("Tokenizer verified in previous run (cached), skipping verification.");
+            tracing::info!(reason = "cached", "skipping server tokenizer verification");
        } else {
            let num_special =
                tokenizer.as_ref().map(|t| t.num_special_tokens_to_add()).unwrap_or(0);
@@ -723,14 +739,17 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
            .await?
            {
                SampleVerifyOutcome::Passed => {
-                    println!("Sample verification passed, skipping full verification.");
+                    tracing::info!("tokenizer sample verification passed");
                    mark_tokenizer_verified(&cache_key);
                }
                SampleVerifyOutcome::Skipped(reason) => {
-                    println!("Server /tokenize unavailable ({reason}), skipping verification.");
+                    tracing::warn!(
+                        reason = %reason,
+                        "server tokenizer unavailable; skipping prompt verification"
+                    );
                }
                SampleVerifyOutcome::Mismatch => {
-                    println!("Sample verification found mismatch, running full verify+fix...");
+                    tracing::warn!("tokenizer sample mismatch; verifying and fixing all prompts");
                    match verify_and_fix_prompt_lengths(
                        &client,
                        &config.base_url,
@@ -742,16 +761,16 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
                    .await
                    {
                        Ok(()) => {
-                            println!(
-                                "All {} prompts verified: exact token length match.",
-                                input_requests.len()
+                            tracing::info!(
+                                prompts = input_requests.len(),
+                                "verified exact prompt token lengths"
                            );
                            mark_tokenizer_verified(&cache_key);
                        }
                        Err(BenchError::TokenizeUnavailable(reason)) => {
-                            println!(
-                                "Server /tokenize became unavailable during verification \
-                                 ({reason}); proceeding with client-side token counts."
+                            tracing::warn!(
+                                reason = %reason,
+                                "server tokenizer became unavailable; using client token counts"
                            );
                        }
                        Err(e) => return Err(e),
@@ -763,7 +782,7 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {

    // Warmup
    if config.num_warmups > 0 {
-        println!("Warming up with {} requests...", config.num_warmups);
+        tracing::info!(requests = config.num_warmups, "starting benchmark warmup");
        run_warmup(
            config.backend,
            &client,
@@ -776,7 +795,7 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
            config.disable_tqdm,
        )
        .await;
-        println!("Warmup run completed.");
+        tracing::info!(requests = config.num_warmups, "benchmark warmup completed");
    }

    // Start profiler if requested (immediate mode — no batch threshold)
@@ -814,28 +833,22 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
    let spec_decode_before =
        fetch_spec_decode_metrics(&config.base_url, &client, &config.extra_headers).await;
    if spec_decode_before.is_some() {
-        println!("Speculative decoding detected, will collect metrics.");
+        tracing::info!("detected speculative decoding; collecting metrics");
    }

    // Main benchmark
-    println!("Starting main benchmark run...");
    let distribution = if config.burstiness == 1.0 {
        "Poisson process"
    } else {
        "Gamma distribution"
    };
-    println!(
-        "Traffic request rate: {}",
-        if config.request_rate.is_infinite() {
-            "inf".to_string()
-        } else {
-            format!("{}", config.request_rate)
-        }
-    );
-    println!("Burstiness factor: {} ({distribution})", config.burstiness);
-    println!(
-        "Maximum request concurrency: {}",
-        config.max_concurrency.unwrap_or(config.num_prompts)
+    tracing::info!(
+        request_rate = config.request_rate,
+        burstiness = config.burstiness,
+        distribution,
+        max_concurrency = config.max_concurrency.unwrap_or(config.num_prompts),
+        prompts = config.num_prompts,
+        "starting main benchmark run"
    );

    // Pre-assign LoRA adapters to each request (None when --lora-modules not set).
@@ -847,11 +860,11 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
    );
    if let (Some(modules), Some(_)) = (config.lora_modules.as_ref(), lora_assignments.as_ref()) {
        let names: Vec<&str> = modules.iter().map(|s| s.as_ref()).collect();
-        println!(
-            "LoRA adapters ({}): {:?} [assignment={:?}]",
-            modules.len(),
-            names,
-            config.lora_assignment
+        tracing::info!(
+            adapters = modules.len(),
+            names = ?names,
+            assignment = ?config.lora_assignment,
+            "assigned LoRA adapters"
        );
    }

@@ -1125,7 +1138,7 @@ pub async fn run_benchmark(config: &BenchConfig) -> Result<serde_json::Value> {
    if let Some((cancel_tx, task)) = profile_task {
        let _ = cancel_tx.send(());
        if let Err(e) = task.await {
-            eprintln!("WARNING: Profile background task failed: {e}");
+            tracing::error!(error = %e.as_report(), "profiler background task failed");
        }
    }

@@ -1289,12 +1302,14 @@ pub(crate) async fn start_profiler_immediate(
    base_url: &str,
    extra_headers: &Option<std::collections::HashMap<String, String>>,
 ) {
-    println!("Starting profiler...");
    let profile_url = format!("{base_url}/start_profile");
+    tracing::info!(url = %profile_url, "starting profiler");
    match send_profile_request(client, &profile_url, extra_headers).await {
-        Ok(true) => println!("Profiler started"),
-        Ok(false) => eprintln!("WARNING: Profiler start request returned non-success"),
-        Err(e) => eprintln!("WARNING: Failed to start profiler: {e}"),
+        Ok(true) => tracing::info!(url = %profile_url, "profiler started"),
+        Ok(false) => tracing::warn!(url = %profile_url, "profiler start request was unsuccessful"),
+        Err(e) => {
+            tracing::warn!(url = %profile_url, error = %e.as_report(), "failed to start profiler")
+        }
    }
 }

@@ -1304,12 +1319,14 @@ pub(crate) async fn stop_profiler_immediate(
    base_url: &str,
    extra_headers: &Option<std::collections::HashMap<String, String>>,
 ) {
-    println!("Stopping profiler...");
    let profile_url = format!("{base_url}/stop_profile");
+    tracing::info!(url = %profile_url, "stopping profiler");
    match send_profile_request(client, &profile_url, extra_headers).await {
-        Ok(true) => println!("Profiler stopped"),
-        Ok(false) => eprintln!("WARNING: Profiler stop request returned non-success"),
-        Err(e) => eprintln!("WARNING: Failed to stop profiler: {e}"),
+        Ok(true) => tracing::info!(url = %profile_url, "profiler stopped"),
+        Ok(false) => tracing::warn!(url = %profile_url, "profiler stop request was unsuccessful"),
+        Err(e) => {
+            tracing::warn!(url = %profile_url, error = %e.as_report(), "failed to stop profiler")
+        }
    }
 }

@@ -1371,25 +1388,30 @@ pub(crate) async fn profile_on_batch_threshold(
    duration_secs: f64,
    mut cancel_rx: tokio::sync::oneshot::Receiver<()>,
 ) {
-    println!(
-        "Waiting for batch size >= {threshold} before starting profiler \
-         (will capture {duration_secs}s)..."
+    tracing::info!(
+        threshold,
+        duration_seconds = duration_secs,
+        "waiting for profiler batch threshold"
    );

    loop {
        if let Some(running) = fetch_num_requests_running(client, base_url).await
            && running >= threshold
        {
-            println!("Batch size {running} >= {threshold}, starting profiler...");
+            tracing::info!(
+                running_requests = running,
+                threshold,
+                "profiler batch threshold reached"
+            );
            break;
        }
        // Wait 500ms or until the benchmark signals cancellation
        tokio::select! {
            _ = tokio::time::sleep(std::time::Duration::from_millis(500)) => {}
            _ = &mut cancel_rx => {
-                eprintln!(
-                    "NOTE: Benchmark finished before batch threshold {threshold} was reached; \
-                     profiling skipped."
+                tracing::warn!(
+                    threshold,
+                    "benchmark finished before profiler batch threshold; skipping profiling"
                );
                return;
            }
@@ -1398,13 +1420,13 @@ pub(crate) async fn profile_on_batch_threshold(

    let start_url = format!("{base_url}/start_profile");
    match send_profile_request(client, &start_url, extra_headers).await {
-        Ok(true) => println!("Profiler started"),
+        Ok(true) => tracing::info!(url = %start_url, "profiler started"),
        Ok(false) => {
-            eprintln!("WARNING: Profiler start request returned non-success");
+            tracing::warn!(url = %start_url, "profiler start request was unsuccessful");
            return;
        }
        Err(e) => {
-            eprintln!("WARNING: Failed to start profiler: {e}");
+            tracing::warn!(url = %start_url, error = %e.as_report(), "failed to start profiler");
            return;
        }
    }
@@ -1413,15 +1435,17 @@ pub(crate) async fn profile_on_batch_threshold(
    tokio::select! {
        _ = tokio::time::sleep(std::time::Duration::from_secs_f64(duration_secs)) => {}
        _ = &mut cancel_rx => {
-            println!("Benchmark finished, stopping profiler early...");
+            tracing::info!("benchmark finished; stopping profiler early");
        }
    }

    let stop_url = format!("{base_url}/stop_profile");
    match send_profile_request(client, &stop_url, extra_headers).await {
-        Ok(true) => println!("Profiler stopped after capturing"),
-        Ok(false) => eprintln!("WARNING: Profiler stop request returned non-success"),
-        Err(e) => eprintln!("WARNING: Failed to stop profiler: {e}"),
+        Ok(true) => tracing::info!(url = %stop_url, "profiler stopped after capture"),
+        Ok(false) => tracing::warn!(url = %stop_url, "profiler stop request was unsuccessful"),
+        Err(e) => {
+            tracing::warn!(url = %stop_url, error = %e.as_report(), "failed to stop profiler")
+        }
    }
 }

@@ -1502,10 +1526,11 @@ async fn verify_and_fix_prompt_lengths(
                let excess = tokens.len().saturating_sub(expected_input_len);
                let compensate = if excess > 0 && last_excess == Some(excess) {
                    if _iter == 1 {
-                        eprintln!(
-                            "Prompt {i}: server consistently adds {excess} extra token(s) \
-                             (likely BOS), compensating target to {}.",
-                            expected_input_len.saturating_sub(excess),
+                        tracing::warn!(
+                            prompt_index = i,
+                            extra_tokens = excess,
+                            adjusted_target = expected_input_len.saturating_sub(excess),
+                            "server consistently adds prompt tokens; compensating verification target"
                        );
                    }
                    excess
@@ -1563,7 +1588,10 @@ async fn verify_and_fix_prompt_lengths(

    let fc = fixed_count.load(std::sync::atomic::Ordering::Relaxed);
    if fc > 0 {
-        println!("Fixed {fc} prompt(s) via server tokenize/detokenize convergence.");
+        tracing::info!(
+            fixed_prompts = fc,
+            "fixed prompt lengths using server tokenizer"
+        );
    }

    Ok(())
@@ -1818,7 +1846,7 @@ async fn sample_verify_prompts(
    let tokenize_url = format!("{base_url}/tokenize");
    let api_key = std::env::var("OPENAI_API_KEY").ok();

-    println!("Sampling {sample_size} prompts for verification...");
+    tracing::info!(sample_size, "sampling prompts for tokenizer verification");

    for (i, request) in requests.iter().enumerate().take(sample_size) {
        let tokens = match server_tokenize(
@@ -1841,9 +1869,11 @@ async fn sample_verify_prompts(

        let expected = request.prompt_len + num_special;
        if tokens.len() != expected {
-            println!(
-                "Prompt {i}: expected {expected} tokens, server returned {}",
-                tokens.len()
+            tracing::warn!(
+                prompt_index = i,
+                expected_tokens = expected,
+                actual_tokens = tokens.len(),
+                "tokenizer verification sample mismatch"
            );
            return Ok(SampleVerifyOutcome::Mismatch);
        }
@@ -288,10 +288,18 @@ impl BenchConfig {
                    }
                    Some(other) => {
                        // extra_body was not an object — just use sampling params
-                        eprintln!(
-                            "Warning: --extra-body is not a JSON object, sampling params may be lost"
+                        let value_type = match &other {
+                            serde_json::Value::Null => "null",
+                            serde_json::Value::Bool(_) => "boolean",
+                            serde_json::Value::Number(_) => "number",
+                            serde_json::Value::String(_) => "string",
+                            serde_json::Value::Array(_) => "array",
+                            serde_json::Value::Object(_) => unreachable!(),
+                        };
+                        tracing::warn!(
+                            value_type,
+                            "sampling parameters may be lost because --extra-body is not a JSON object"
                        );
-                        let _ = other;
                        sampling_params
                    }
                    None => sampling_params,
@@ -489,9 +497,9 @@ impl BenchConfig {
                _ => {}
            }
            if !args.skip_chat_template {
-                eprintln!(
-                    "NOTE: client-side chat template rendering is not supported; custom \
-                     dataset prompts are sent raw (equivalent to --skip-chat-template)."
+                tracing::warn!(
+                    dataset = "custom",
+                    "client-side chat template rendering is unsupported; sending prompts raw"
                );
            }
        }
@@ -570,9 +578,10 @@ impl BenchConfig {
            }

            if ignore_eos {
-                eprintln!(
-                    "WARNING: --ignore-eos is set with --multi-turn. The server may not \
-                     respect output length limits, causing unbounded context growth."
+                tracing::warn!(
+                    ignore_eos,
+                    multi_turn = true,
+                    "output length limits may be ignored, causing unbounded context growth"
                );
            }

@@ -128,8 +128,10 @@ mod tests {

    /// gpt2 via built-in tiktoken encoding — loads without network access.
    fn test_tokenizer() -> TokenizerKind {
-        crate::tokenizer::load_tokenizer("gpt2", false, None)
-            .expect("gpt2 built-in tiktoken should always load without network")
+        TokenizerKind::Tiktoken(
+            crate::tiktoken::load_builtin_tiktoken("gpt2")
+                .expect("gpt2 built-in tiktoken should always load without network"),
+        )
    }

    #[test]
@@ -8,6 +8,7 @@ use rand::seq::SliceRandom;
 use rand::{Rng, SeedableRng};

 use super::SampleRequest;
+use super::progress::RowDownloadReporter;
 use crate::error::{BenchError, Result};
 use crate::tokenizer::TokenizerKind;

@@ -50,18 +51,19 @@ enum ColumnFormat {

 /// Make a GET request with retry logic (3 retries with exponential backoff).
 /// Returns the parsed JSON response.
-fn get_with_retry(
-    client: &reqwest::blocking::Client,
+async fn get_with_retry(
+    client: &reqwest::Client,
    url: &str,
    label: &str,
 ) -> Result<serde_json::Value> {
    let max_retries = 3;
    for attempt in 0..=max_retries {
-        let resp = match client.get(url).send() {
+        let resp = match client.get(url).send().await {
            Ok(r) => r,
            Err(e) => {
                if attempt < max_retries {
-                    std::thread::sleep(std::time::Duration::from_secs(2 * (attempt as u64 + 1)));
+                    tokio::time::sleep(std::time::Duration::from_secs(2 * (attempt as u64 + 1)))
+                        .await;
                    continue;
                }
                return Err(BenchError::Config(format!(
@@ -80,7 +82,7 @@ fn get_with_retry(
        }

        if status.is_server_error() && attempt < max_retries {
-            std::thread::sleep(std::time::Duration::from_secs(2 * (attempt as u64 + 1)));
+            tokio::time::sleep(std::time::Duration::from_secs(2 * (attempt as u64 + 1))).await;
            continue;
        }

@@ -92,6 +94,7 @@ fn get_with_retry(

        let data: serde_json::Value = resp
            .json()
+            .await
            .map_err(|e| BenchError::Config(format!("Failed to parse {label} response: {e}")))?;
        return Ok(data);
    }
@@ -105,7 +108,7 @@ fn get_with_retry(
 /// If both `subset` and `split` are provided, the `/info` call is skipped as an optimization.
 /// Paginated download fetches rows in pages of 100 until `num_rows_needed` are collected
 /// or the dataset is exhausted.
-pub fn download_hf_dataset(
+pub async fn download_hf_dataset(
    dataset: &str,
    subset: Option<&str>,
    split: Option<&str>,
@@ -115,7 +118,7 @@ pub fn download_hf_dataset(
        url::form_urlencoded::byte_serialize(dataset.as_bytes()).collect();

    let mut client_builder =
-        reqwest::blocking::Client::builder().timeout(std::time::Duration::from_secs(120));
+        reqwest::Client::builder().timeout(std::time::Duration::from_secs(120));

    // Add HF_TOKEN auth header if available
    if let Ok(token) = std::env::var("HF_TOKEN") {
@@ -138,7 +141,7 @@ pub fn download_hf_dataset(
        // Call /info to discover available configs and splits
        let info_url =
            format!("https://datasets-server.huggingface.co/info?dataset={encoded_dataset}");
-        let info = get_with_retry(&client, &info_url, "HF dataset /info")?;
+        let info = get_with_retry(&client, &info_url, "HF dataset /info").await?;

        let dataset_info =
            info.get("dataset_info").and_then(|d| d.as_object()).ok_or_else(|| {
@@ -201,7 +204,12 @@ pub fn download_hf_dataset(
        (resolved_config, resolved_split)
    };

-    println!("HF dataset: {dataset} (config={resolved_config}, split={resolved_split})");
+    tracing::info!(
+        dataset,
+        config = resolved_config,
+        split = resolved_split,
+        "resolved Hugging Face dataset"
+    );

    // Check cache
    let dir = cache_dir();
@@ -215,11 +223,16 @@ pub fn download_hf_dataset(

    if cache_path.exists() {
        let path_str = cache_path.to_string_lossy().to_string();
-        println!("HF dataset cached: {path_str}");
+        tracing::info!(dataset, path = %path_str, "using cached Hugging Face dataset");
        return Ok((path_str, resolved_config, resolved_split));
    }

-    println!("Downloading HF dataset '{dataset}' from datasets-server...");
+    tracing::info!(
+        dataset,
+        config = resolved_config,
+        split = resolved_split,
+        "downloading Hugging Face dataset"
+    );

    let encoded_config: String =
        url::form_urlencoded::byte_serialize(resolved_config.as_bytes()).collect();
@@ -229,6 +242,7 @@ pub fn download_hf_dataset(
    let mut all_rows: Vec<serde_json::Value> = Vec::new();
    let mut offset = 0usize;
    let page_size = 100usize;
+    let mut progress = RowDownloadReporter::new();

    loop {
        let url = format!(
@@ -240,7 +254,7 @@ pub fn download_hf_dataset(
             &length={page_size}"
        );

-        let data = get_with_retry(&client, &url, "HF dataset /rows")?;
+        let data = get_with_retry(&client, &url, "HF dataset /rows").await?;

        let rows = data["rows"]
            .as_array()
@@ -260,14 +274,14 @@ pub fn download_hf_dataset(
        offset += fetched;

        let total = data["num_rows_total"].as_u64().unwrap_or(0);
-        eprint!("\r  Fetched {offset}/{total} rows...");
+        progress.update(offset, total);

        // Stop if we have enough rows or reached end of dataset
        if all_rows.len() >= num_rows_needed || fetched < page_size {
            break;
        }
    }
-    eprintln!(); // newline after progress
+    progress.finish();

    if all_rows.is_empty() {
        return Err(BenchError::Config(format!(
@@ -280,7 +294,12 @@ pub fn download_hf_dataset(
    std::fs::write(&cache_path, &json_str)?;

    let path_str = cache_path.to_string_lossy().to_string();
-    println!("HF dataset: {} rows saved to {path_str}", all_rows.len());
+    tracing::info!(
+        dataset,
+        rows = all_rows.len(),
+        path = %path_str,
+        "saved Hugging Face dataset"
+    );
    Ok((path_str, resolved_config, resolved_split))
 }

@@ -483,21 +502,31 @@ pub fn load_hf_dataset(
    // Detect column format from first row
    let format = detect_column_format(&entries[0], text_column_override)?;

-    // Print detected format
    match &format {
-        ColumnFormat::Chat(col) => println!("HF dataset: detected chat column '{col}'"),
+        ColumnFormat::Chat(col) => {
+            tracing::info!(
+                format = "chat",
+                column = col,
+                "detected Hugging Face dataset format"
+            );
+        }
        ColumnFormat::Text {
            prompt_col,
            output_col,
        } => {
-            let out_msg = output_col.as_deref().unwrap_or("none");
-            println!("HF dataset: detected text column '{prompt_col}', output column: {out_msg}");
+            tracing::info!(
+                format = "text",
+                prompt_column = prompt_col,
+                output_column = output_col.as_deref().unwrap_or("none"),
+                "detected Hugging Face dataset format"
+            );
        }
        ColumnFormat::Combined { cols, output_col } => {
-            let out_msg = output_col.as_deref().unwrap_or("none");
-            println!(
-                "HF dataset: detected combined columns {:?}, output column: {out_msg}",
-                cols
+            tracing::info!(
+                format = "combined",
+                prompt_columns = ?cols,
+                output_column = output_col.as_deref().unwrap_or("none"),
+                "detected Hugging Face dataset format"
            );
        }
    }
@@ -608,9 +637,10 @@ pub fn load_hf_dataset(
                if len == 0 { 128 } else { len }
            } else {
                if !warned_no_output {
-                    eprintln!(
-                        "WARNING: No output column detected and --hf-output-len not set. \
-                         Using default output length of 128 tokens."
+                    tracing::warn!(
+                        path = dataset_path,
+                        default_output_tokens = 128,
+                        "no dataset output column or --hf-output-len; using default output length"
                    );
                    warned_no_output = true;
                }
@@ -618,9 +648,10 @@ pub fn load_hf_dataset(
            }
        } else {
            if !warned_no_output {
-                eprintln!(
-                    "WARNING: No output column detected and --hf-output-len not set. \
-                     Using default output length of 128 tokens."
+                tracing::warn!(
+                    path = dataset_path,
+                    default_output_tokens = 128,
+                    "no dataset output column or --hf-output-len; using default output length"
                );
                warned_no_output = true;
            }
@@ -640,9 +671,11 @@ pub fn load_hf_dataset(
    // Oversample if needed
    if samples.len() < num_requests {
        if no_oversample {
-            println!(
-                "Skipping oversampling. Total samples: {} (requested: {num_requests})",
-                samples.len()
+            tracing::info!(
+                dataset = "hf",
+                samples = samples.len(),
+                requested = num_requests,
+                "skipping dataset oversampling"
            );
        } else if !samples.is_empty() {
            let original_len = samples.len();
@@ -652,9 +685,11 @@ pub fn load_hf_dataset(
                req.request_id = Some(format!("{request_id_prefix}{}", original_len + i));
                samples.push(req);
            }
-            println!(
-                "Oversampled HF dataset from {original_len} to {} total samples.",
-                samples.len()
+            tracing::info!(
+                dataset = "hf",
+                original_samples = original_len,
+                samples = samples.len(),
+                "oversampled dataset"
            );
        }
    }
@@ -1002,8 +1037,10 @@ mod tests {

    /// Build a gpt2 tokenizer using built-in tiktoken encoding (no network required).
    fn builtin_tokenizer() -> crate::tokenizer::TokenizerKind {
-        crate::tokenizer::load_tokenizer("gpt2", false, None)
-            .expect("gpt2 built-in tiktoken should always load without network")
+        crate::tokenizer::TokenizerKind::Tiktoken(
+            crate::tiktoken::load_builtin_tiktoken("gpt2")
+                .expect("gpt2 built-in tiktoken should always load without network"),
+        )
    }

    /// Write JSON data to a unique temp file and return the path string.
@@ -5,6 +5,7 @@ pub mod custom;
 pub mod hf_dataset;
 pub mod multi_turn;
 pub mod prefix_repetition;
+mod progress;
 pub mod random;
 pub mod random_mm;
 pub mod random_rerank;
@@ -90,9 +91,10 @@ pub fn oversample_requests(
        return;
    }
    if no_oversample {
-        println!(
-            "Skipping oversampling. Total samples: {} (requested: {num_requests})",
-            requests.len()
+        tracing::info!(
+            samples = requests.len(),
+            requested = num_requests,
+            "skipping dataset oversampling"
        );
        return;
    }
@@ -103,9 +105,10 @@ pub fn oversample_requests(
        req.request_id = Some(format!("{request_id_prefix}{}", original_len + i));
        requests.push(req);
    }
-    println!(
-        "Oversampled requests from {original_len} to {} total samples.",
-        requests.len()
+    tracing::info!(
+        original_samples = original_len,
+        samples = requests.len(),
+        "oversampled dataset"
    );
 }

@@ -445,9 +445,10 @@ pub fn load_sharegpt_multi_turn(
            conv.conversation_id = format!("{request_id_prefix}conv-{}", original_len + i);
            conversations.push(conv);
        }
-        println!(
-            "Oversampled multi-turn conversations from {original_len} to {} total.",
-            conversations.len()
+        tracing::info!(
+            original_conversations = original_len,
+            conversations = conversations.len(),
+            "oversampled multi-turn conversations"
        );
    }

@@ -525,10 +526,12 @@ mod tests {
        len
    }

-    #[test]
+    #[tokio::test]
    #[ignore]
-    fn test_prefix_sharing_structure() {
-        let tok = crate::tokenizer::load_tokenizer("nvidia/Kimi-K2.5-NVFP4", false, None).unwrap();
+    async fn test_prefix_sharing_structure() {
+        let tok = crate::tokenizer::load_tokenizer("nvidia/Kimi-K2.5-NVFP4", false, None)
+            .await
+            .unwrap();

        let cfg = MultiTurnRandomConfig {
            num_conversations: 5,
@@ -610,10 +613,12 @@ mod tests {
        println!("All prefix sharing checks passed!");
    }

-    #[test]
+    #[tokio::test]
    #[ignore]
-    fn test_per_turn_input_len_default_mode() {
-        let tok = crate::tokenizer::load_tokenizer("nvidia/Kimi-K2.5-NVFP4", false, None).unwrap();
+    async fn test_per_turn_input_len_default_mode() {
+        let tok = crate::tokenizer::load_tokenizer("nvidia/Kimi-K2.5-NVFP4", false, None)
+            .await
+            .unwrap();

        let cfg = MultiTurnRandomConfig {
            num_conversations: 4,
@@ -650,10 +655,12 @@ mod tests {
        println!("per_turn_input_len default-mode checks passed!");
    }

-    #[test]
+    #[tokio::test]
    #[ignore]
-    fn test_variable_turns_range() {
-        let tok = crate::tokenizer::load_tokenizer("nvidia/Kimi-K2.5-NVFP4", false, None).unwrap();
+    async fn test_variable_turns_range() {
+        let tok = crate::tokenizer::load_tokenizer("nvidia/Kimi-K2.5-NVFP4", false, None)
+            .await
+            .unwrap();

        let cfg = MultiTurnRandomConfig {
            num_conversations: 50,
@@ -684,10 +691,12 @@ mod tests {
        println!("variable_turns_range checks passed! counts: {distinct_counts:?}");
    }

-    #[test]
+    #[tokio::test]
    #[ignore]
-    fn test_variable_turns_fixed() {
-        let tok = crate::tokenizer::load_tokenizer("nvidia/Kimi-K2.5-NVFP4", false, None).unwrap();
+    async fn test_variable_turns_fixed() {
+        let tok = crate::tokenizer::load_tokenizer("nvidia/Kimi-K2.5-NVFP4", false, None)
+            .await
+            .unwrap();

        let cfg = MultiTurnRandomConfig {
            num_conversations: 10,
@@ -709,10 +718,12 @@ mod tests {
        println!("variable_turns_fixed checks passed!");
    }

-    #[test]
+    #[tokio::test]
    #[ignore]
-    fn test_per_turn_input_len_prefix_sharing() {
-        let tok = crate::tokenizer::load_tokenizer("nvidia/Kimi-K2.5-NVFP4", false, None).unwrap();
+    async fn test_per_turn_input_len_prefix_sharing() {
+        let tok = crate::tokenizer::load_tokenizer("nvidia/Kimi-K2.5-NVFP4", false, None)
+            .await
+            .unwrap();

        // Turn 0 input_len=1000, turns 1+ per_turn_input_len=600
        // global_len ≈ 100 (10%), conv_len ≈ 800 (80%), unique ≈ 100
@@ -41,11 +41,13 @@ pub fn generate_prefix_repetition_dataset(
    }
    let total = prompts_per_prefix * num_prefixes;
    if total != num_requests {
-        println!(
-            "prefix_repetition: generating {total} requests \
-             ({num_prefixes} prefixes x {prompts_per_prefix} prompts each; \
-             {} dropped to divide evenly)",
-            num_requests - total
+        tracing::info!(
+            requested = num_requests,
+            generated = total,
+            prefixes = num_prefixes,
+            prompts_per_prefix,
+            dropped = num_requests - total,
+            "adjusted prefix-repetition request count"
        );
    }

@@ -109,8 +111,10 @@ mod tests {

    /// gpt2 via built-in tiktoken encoding — loads without network access.
    fn test_tokenizer() -> TokenizerKind {
-        crate::tokenizer::load_tokenizer("gpt2", false, None)
-            .expect("gpt2 built-in tiktoken should always load without network")
+        TokenizerKind::Tiktoken(
+            crate::tiktoken::load_builtin_tiktoken("gpt2")
+                .expect("gpt2 built-in tiktoken should always load without network"),
+        )
    }

    #[test]
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+use std::time::{Duration, Instant};
+
+use indicatif::{ProgressBar, ProgressStyle};
+
+const REPORT_INTERVAL: Duration = Duration::from_secs(10);
+
+/// Reports row download progress to an interactive progress bar, or through
+/// periodic tracing events when the progress bar is hidden on a non-TTY.
+pub(super) struct RowDownloadReporter {
+    progress: ProgressBar,
+    next_report: Instant,
+}
+
+impl RowDownloadReporter {
+    /// Creates a reporter that emits non-TTY updates every 10 seconds.
+    pub fn new() -> Self {
+        let progress = ProgressBar::new(0);
+        progress.set_style(
+            ProgressStyle::with_template(
+                "{spinner:.green} Fetching rows [{bar:30.cyan/blue}] {pos}/{len}",
+            )
+            .unwrap()
+            .progress_chars("#>-"),
+        );
+        Self {
+            progress,
+            next_report: Instant::now() + REPORT_INTERVAL,
+        }
+    }
+
+    /// Updates the current row count and reports progress when due.
+    pub fn update(&mut self, rows: usize, total: u64) {
+        let rows = rows as u64;
+        let total = total.max(rows);
+        self.progress.set_length(total);
+        self.progress.set_position(rows);
+
+        if self.should_report(Instant::now()) {
+            tracing::info!(rows, total, "fetching dataset rows");
+        }
+    }
+
+    /// Clears the interactive progress bar after the download completes.
+    pub fn finish(self) {
+        self.progress.finish_and_clear();
+    }
+
+    fn should_report(&mut self, now: Instant) -> bool {
+        if !self.progress.is_hidden() || now < self.next_report {
+            return false;
+        }
+        self.next_report = now + REPORT_INTERVAL;
+        true
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn hidden_reporter_uses_ten_second_deadline() {
+        let start = Instant::now();
+        let mut reporter = RowDownloadReporter {
+            progress: ProgressBar::hidden(),
+            next_report: start + REPORT_INTERVAL,
+        };
+
+        assert!(!reporter.should_report(start + Duration::from_secs(9)));
+        assert!(reporter.should_report(start + Duration::from_secs(10)));
+        assert!(!reporter.should_report(start + Duration::from_secs(19)));
+        assert!(reporter.should_report(start + Duration::from_secs(20)));
+    }
+}
@@ -49,9 +49,12 @@ pub fn generate_random_dataset(
    let (input_low, input_high) = range_ratio.input_bounds(real_input_len);
    let (output_low, output_high) = range_ratio.output_bounds(output_len);
    if !range_ratio.is_fixed() {
-        println!(
-            "Sampling input_len from [{input_low}, {input_high}] and \
-             output_len from [{output_low}, {output_high}]"
+        tracing::info!(
+            input_low,
+            input_high,
+            output_low,
+            output_high,
+            "sampling random request lengths"
        );
    }

@@ -305,7 +308,8 @@ mod tests {
    #[test]
    #[ignore]
    fn test_generate_random_dataset_token_ids() {
-        let tokenizer = tokenizer::load_tokenizer("gpt2", false, None).unwrap();
+        let tokenizer =
+            TokenizerKind::Tiktoken(crate::tiktoken::load_builtin_tiktoken("gpt2").unwrap());
        let requests = generate_random_dataset(
            &tokenizer,
            10,  // num_requests
@@ -337,7 +341,8 @@ mod tests {
    #[test]
    #[ignore]
    fn test_generate_random_dataset_text() {
-        let tokenizer = tokenizer::load_tokenizer("gpt2", false, None).unwrap();
+        let tokenizer =
+            TokenizerKind::Tiktoken(crate::tiktoken::load_builtin_tiktoken("gpt2").unwrap());
        let requests = generate_random_dataset(
            &tokenizer,
            10,  // num_requests
@@ -371,7 +376,8 @@ mod tests {
    #[test]
    #[ignore]
    fn test_token_length_exact_local() {
-        let tokenizer = tokenizer::load_tokenizer("gpt2", false, None).unwrap();
+        let tokenizer =
+            TokenizerKind::Tiktoken(crate::tiktoken::load_builtin_tiktoken("gpt2").unwrap());
        let target_len = 512;
        let requests = generate_random_dataset(
            &tokenizer,
@@ -405,11 +411,11 @@ mod tests {
    }

    /// Test that tiktoken tokenizer produces exact target token lengths (token ID mode).
-    #[test]
+    #[tokio::test]
    #[ignore]
-    fn test_token_length_exact_tiktoken() {
+    async fn test_token_length_exact_tiktoken() {
        // Use Qwen2.5 which has a tiktoken-format tokenizer
-        let tokenizer = tokenizer::load_tokenizer("Qwen/Qwen2.5-0.5B", false, None);
+        let tokenizer = tokenizer::load_tokenizer("Qwen/Qwen2.5-0.5B", false, None).await;
        let tokenizer = match tokenizer {
            Ok(t) => t,
            Err(e) => {
@@ -453,10 +459,10 @@ mod tests {

    /// Test encode/decode roundtrip stability for tiktoken.
    /// After one decode→encode cycle with UTF-8-safe tokens, length must not drift.
-    #[test]
+    #[tokio::test]
    #[ignore]
-    fn test_tiktoken_roundtrip_stability() {
-        let tokenizer = tokenizer::load_tokenizer("Qwen/Qwen2.5-0.5B", false, None);
+    async fn test_tiktoken_roundtrip_stability() {
+        let tokenizer = tokenizer::load_tokenizer("Qwen/Qwen2.5-0.5B", false, None).await;
        let tokenizer = match tokenizer {
            Ok(t) => t,
            Err(e) => {
@@ -139,8 +139,10 @@ mod tests {

    /// gpt2 via built-in tiktoken encoding — loads without network access.
    fn test_tokenizer() -> TokenizerKind {
-        crate::tokenizer::load_tokenizer("gpt2", false, None)
-            .expect("gpt2 built-in tiktoken should always load without network")
+        TokenizerKind::Tiktoken(
+            crate::tiktoken::load_builtin_tiktoken("gpt2")
+                .expect("gpt2 built-in tiktoken should always load without network"),
+        )
    }

    fn fixed_ratio() -> RangeRatio {
@@ -22,18 +22,21 @@ const DEFAULT_SHAREGPT_FILE: &str = "ShareGPT_V3_unfiltered_cleaned_split.json";

 /// Download the default ShareGPT dataset from HuggingFace Hub.
 /// Uses hf-hub's built-in cache — subsequent calls return the cached path instantly.
-pub fn download_sharegpt_dataset() -> Result<String> {
-    println!(
-        "Downloading ShareGPT dataset from {DEFAULT_SHAREGPT_REPO}/{DEFAULT_SHAREGPT_FILE} ..."
+pub async fn download_sharegpt_dataset() -> Result<String> {
+    tracing::info!(
+        repository = DEFAULT_SHAREGPT_REPO,
+        file = DEFAULT_SHAREGPT_FILE,
+        "downloading ShareGPT dataset"
    );
-    let repo = crate::hub::HubRepo::dataset(DEFAULT_SHAREGPT_REPO.to_string());
-    let path = repo.get(DEFAULT_SHAREGPT_FILE).map_err(|e| {
+    let repo = crate::hub::HubRepo::dataset(DEFAULT_SHAREGPT_REPO.to_string())
+        .map_err(BenchError::Config)?;
+    let path = repo.get(DEFAULT_SHAREGPT_FILE).await.map_err(|e| {
        BenchError::Config(format!(
            "Failed to download ShareGPT dataset from '{DEFAULT_SHAREGPT_REPO}': {e}"
        ))
    })?;
    let path_str = path.to_string_lossy().to_string();
-    println!("ShareGPT dataset ready: {path_str}");
+    tracing::info!(dataset = "sharegpt", path = %path_str, "dataset is ready");
    Ok(path_str)
 }

@@ -135,9 +138,11 @@ pub fn load_sharegpt_dataset(
    // Oversample if dataset is smaller than requested
    if samples.len() < num_requests {
        if no_oversample {
-            println!(
-                "Skipping oversampling. Total samples: {} (requested: {num_requests})",
-                samples.len()
+            tracing::info!(
+                dataset = "sharegpt",
+                samples = samples.len(),
+                requested = num_requests,
+                "skipping dataset oversampling"
            );
        } else if !samples.is_empty() {
            let needed = num_requests - samples.len();
@@ -147,9 +152,11 @@ pub fn load_sharegpt_dataset(
                req.request_id = Some(format!("{request_id_prefix}{}", original_len + i));
                samples.push(req);
            }
-            println!(
-                "Oversampled requests from {original_len} to {} total samples.",
-                samples.len()
+            tracing::info!(
+                dataset = "sharegpt",
+                original_samples = original_len,
+                samples = samples.len(),
+                "oversampled dataset"
            );
        }
    }
@@ -8,6 +8,7 @@ use rand::seq::SliceRandom;
 use rand::{Rng, SeedableRng};

 use super::SampleRequest;
+use super::progress::RowDownloadReporter;
 use crate::cli::SpeedBenchConfig;
 use crate::error::{BenchError, Result};
 use crate::tokenizer::TokenizerKind;
@@ -25,7 +26,7 @@ fn cache_dir() -> std::path::PathBuf {

 /// Download SPEED-Bench dataset from HuggingFace datasets-server API.
 /// Results are cached as JSON locally for subsequent runs.
-pub fn download_speed_bench(config: SpeedBenchConfig) -> Result<String> {
+pub async fn download_speed_bench(config: SpeedBenchConfig) -> Result<String> {
    let config_name = config.as_str();

    let dir = cache_dir();
@@ -35,13 +36,13 @@ pub fn download_speed_bench(config: SpeedBenchConfig) -> Result<String> {
    // Return cached file if it exists
    if cache_path.exists() {
        let path_str = cache_path.to_string_lossy().to_string();
-        println!("SPEED-Bench ({config_name}) cached: {path_str}");
+        tracing::info!(config = config_name, path = %path_str, "using cached SPEED-Bench dataset");
        return Ok(path_str);
    }

-    println!("Downloading SPEED-Bench ({config_name}) from HuggingFace datasets-server...");
+    tracing::info!(config = config_name, "downloading SPEED-Bench dataset");

-    let client = reqwest::blocking::Client::builder()
+    let client = reqwest::Client::builder()
        .timeout(std::time::Duration::from_secs(120))
        .build()
        .map_err(|e| BenchError::Config(format!("Failed to build HTTP client: {e}")))?;
@@ -49,6 +50,7 @@ pub fn download_speed_bench(config: SpeedBenchConfig) -> Result<String> {
    let mut all_rows: Vec<serde_json::Value> = Vec::new();
    let mut offset = 0usize;
    let page_size = 100usize;
+    let mut progress = RowDownloadReporter::new();

    loop {
        let url = format!(
@@ -64,13 +66,14 @@ pub fn download_speed_bench(config: SpeedBenchConfig) -> Result<String> {
        let max_retries = 3;
        let mut data: Option<serde_json::Value> = None;
        for attempt in 0..=max_retries {
-            let resp = match client.get(&url).send() {
+            let resp = match client.get(&url).send().await {
                Ok(r) => r,
                Err(e) => {
                    if attempt < max_retries {
-                        std::thread::sleep(std::time::Duration::from_secs(
+                        tokio::time::sleep(std::time::Duration::from_secs(
                            2 * (attempt as u64 + 1),
-                        ));
+                        ))
+                        .await;
                        continue;
                    }
                    return Err(BenchError::Config(format!(
@@ -80,7 +83,7 @@ pub fn download_speed_bench(config: SpeedBenchConfig) -> Result<String> {
            };

            if resp.status().is_server_error() && attempt < max_retries {
-                std::thread::sleep(std::time::Duration::from_secs(2 * (attempt as u64 + 1)));
+                tokio::time::sleep(std::time::Duration::from_secs(2 * (attempt as u64 + 1))).await;
                continue;
            }

@@ -91,7 +94,7 @@ pub fn download_speed_bench(config: SpeedBenchConfig) -> Result<String> {
                )));
            }

-            data = Some(resp.json().map_err(|e| {
+            data = Some(resp.json().await.map_err(|e| {
                BenchError::Config(format!("Failed to parse SPEED-Bench API response: {e}"))
            })?);
            break;
@@ -116,15 +119,14 @@ pub fn download_speed_bench(config: SpeedBenchConfig) -> Result<String> {
        let fetched = rows.len();
        offset += fetched;

-        // Print progress
        let total = data["num_rows_total"].as_u64().unwrap_or(0);
-        eprint!("\r  Fetched {offset}/{total} rows...");
+        progress.update(offset, total);

        if fetched < page_size {
            break;
        }
    }
-    eprintln!(); // newline after progress
+    progress.finish();

    if all_rows.is_empty() {
        return Err(BenchError::Config(
@@ -137,9 +139,11 @@ pub fn download_speed_bench(config: SpeedBenchConfig) -> Result<String> {
    std::fs::write(&cache_path, &json_str)?;

    let path_str = cache_path.to_string_lossy().to_string();
-    println!(
-        "SPEED-Bench ({config_name}): {} rows saved to {path_str}",
-        all_rows.len()
+    tracing::info!(
+        config = config_name,
+        rows = all_rows.len(),
+        path = %path_str,
+        "saved SPEED-Bench dataset"
    );
    Ok(path_str)
 }
@@ -263,9 +267,11 @@ pub fn load_speed_bench_dataset(
    // Oversample if needed
    if samples.len() < num_requests {
        if no_oversample {
-            println!(
-                "Skipping oversampling. Total samples: {} (requested: {num_requests})",
-                samples.len()
+            tracing::info!(
+                dataset = "speed-bench",
+                samples = samples.len(),
+                requested = num_requests,
+                "skipping dataset oversampling"
            );
        } else if !samples.is_empty() {
            let original_len = samples.len();
@@ -275,9 +281,11 @@ pub fn load_speed_bench_dataset(
                req.request_id = Some(format!("{request_id_prefix}{}", original_len + i));
                samples.push(req);
            }
-            println!(
-                "Oversampled SPEED-Bench from {original_len} to {} total samples.",
-                samples.len()
+            tracing::info!(
+                dataset = "speed-bench",
+                original_samples = original_len,
+                samples = samples.len(),
+                "oversampled dataset"
            );
        }
    }
@@ -288,7 +296,6 @@ pub fn load_speed_bench_dataset(
        ));
    }

-    // Print category distribution
    let mut cat_counts: std::collections::HashMap<&str, usize> = std::collections::HashMap::new();
    for entry in &filtered[..filtered.len().min(samples.len())] {
        let cat = entry.get("category").and_then(|c| c.as_str()).unwrap_or("unknown");
@@ -297,7 +304,7 @@ pub fn load_speed_bench_dataset(
    let mut cats: Vec<_> = cat_counts.into_iter().collect();
    cats.sort_by_key(|b| std::cmp::Reverse(b.1));
    let cat_str: Vec<String> = cats.iter().map(|(k, v)| format!("{k}:{v}")).collect();
-    println!("SPEED-Bench categories: {}", cat_str.join(", "));
+    tracing::info!(categories = %cat_str.join(", "), "computed SPEED-Bench category distribution");

    Ok(samples)
 }
@@ -1,54 +1,39 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

-//! Sync facade over the async `hf_hub` API.
-//!
-//! The workspace bans rustls (`rust/deny.toml`), but hf-hub's sync `ureq`
-//! backend unconditionally pulls ureq's default rustls feature. So we use the
-//! reqwest/native-tls tokio API instead, and bridge blocking callers (dataset
-//! loaders, tokenizer fallback in rayon threads) by running each download on a
-//! dedicated thread with its own single-threaded runtime.
-
 use std::path::PathBuf;

+use hf_hub::Repo;
+use hf_hub::api::tokio::{ApiBuilder, ApiRepo};
+
 /// A handle to a HuggingFace Hub repo, downloading via hf-hub's on-disk cache.
 pub struct HubRepo {
-    repo: hf_hub::Repo,
+    repo: ApiRepo,
 }

 impl HubRepo {
-    pub fn model(model_id: String) -> Self {
-        Self {
-            repo: hf_hub::Repo::model(model_id),
-        }
+    pub fn model(model_id: String) -> Result<Self, String> {
+        Self::new(Repo::model(model_id))
    }

-    pub fn dataset(repo_id: String) -> Self {
-        Self {
-            repo: hf_hub::Repo::dataset(repo_id),
+    pub fn dataset(repo_id: String) -> Result<Self, String> {
+        Self::new(Repo::dataset(repo_id))
+    }
+
+    fn new(repo: Repo) -> Result<Self, String> {
+        let mut builder = ApiBuilder::from_env();
+        if let Ok(token) = std::env::var("HF_TOKEN") {
+            builder = builder.with_token(Some(token));
        }
+        let api = builder.build().map_err(|e| format!("Failed to init HF API: {e}"))?;
+        Ok(Self {
+            repo: api.repo(repo),
+        })
    }

    /// Download (or fetch from cache) a single file from the repo.
    /// Auth is handled by hf-hub via HF_TOKEN / the cached login token.
-    pub fn get(&self, filename: &str) -> Result<PathBuf, String> {
-        let repo = self.repo.clone();
-        let filename = filename.to_string();
-        std::thread::spawn(move || {
-            let rt = tokio::runtime::Builder::new_current_thread()
-                .enable_all()
-                .build()
-                .map_err(|e| format!("Failed to build download runtime: {e}"))?;
-            rt.block_on(async move {
-                let mut builder = hf_hub::api::tokio::ApiBuilder::from_env();
-                if let Ok(token) = std::env::var("HF_TOKEN") {
-                    builder = builder.with_token(Some(token));
-                }
-                let api = builder.build().map_err(|e| format!("Failed to init HF API: {e}"))?;
-                api.repo(repo).get(&filename).await.map_err(|e| format!("{e}"))
-            })
-        })
-        .join()
-        .map_err(|_| "HF Hub download thread panicked".to_string())?
+    pub async fn get(&self, filename: &str) -> Result<PathBuf, String> {
+        self.repo.get(filename).await.map_err(|e| format!("{e}"))
    }
 }
@@ -33,7 +33,7 @@ pub fn prepare_process() {
    if let Ok(new) = rlimit::increase_nofile_limit(u64::MAX)
        && new > 1024
    {
-        eprintln!("Open-file limit: {new}");
+        tracing::info!(soft_limit = new, "raised open-file limit");
    }
 }

@@ -19,7 +19,19 @@ struct Cli {
    args: vllm_bench::BenchServeArgs,
 }

+// TODO: unify the tracing subscriber used by different binaries.
+fn init_tracing() {
+    let filter = tracing_subscriber::EnvFilter::try_from_default_env()
+        .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("info"));
+    let _ = tracing_subscriber::fmt()
+        .with_env_filter(filter)
+        .with_writer(std::io::stderr)
+        .try_init();
+}
+
 fn main() -> anyhow::Result<()> {
+    init_tracing();
+
    let cli = Cli::parse();
    vllm_bench::prepare_process();

@@ -7,6 +7,22 @@ use crate::datasets::SampleRequest;
 use crate::metrics::{BenchmarkMetrics, MultiTurnMetrics};
 use crate::multi_turn::ConversationOutput;

+fn log_failed_requests(outputs: &[RequestFuncOutput]) {
+    let failed_outputs: Vec<_> = outputs.iter().filter(|output| !output.success).collect();
+    if failed_outputs.is_empty() {
+        return;
+    }
+
+    tracing::warn!(
+        failed_requests = failed_outputs.len(),
+        displayed_errors = failed_outputs.len().min(10),
+        "benchmark requests failed"
+    );
+    for (index, output) in failed_outputs.into_iter().take(10).enumerate() {
+        tracing::warn!(index, error = %output.error, "benchmark request failed");
+    }
+}
+
 /// Calculate benchmark metrics from request outputs.
 ///
 /// Mirrors Python's `calculate_metrics()` from serve.py:392-599.
@@ -63,14 +79,7 @@ pub fn calculate_metrics(

    let failed = outputs.len() - completed;

-    // Print failed request errors (capped to 10)
-    let failed_outputs: Vec<&RequestFuncOutput> = outputs.iter().filter(|o| !o.success).collect();
-    if !failed_outputs.is_empty() {
-        eprintln!("Failed requests during benchmark run detected (capping to 10):");
-        for (i, err) in failed_outputs.iter().take(10).enumerate() {
-            eprintln!("Error {i}: {}", err.error);
-        }
-    }
+    log_failed_requests(outputs);

    // Calculate max output tokens per second and max concurrent requests
    let mut max_output_tokens_per_s = 0.0_f64;
@@ -295,14 +304,7 @@ pub fn calculate_embedding_metrics(

    let failed = outputs.len() - completed;

-    // Print failed request errors (capped to 10)
-    let failed_outputs: Vec<&RequestFuncOutput> = outputs.iter().filter(|o| !o.success).collect();
-    if !failed_outputs.is_empty() {
-        eprintln!("Failed requests during benchmark run detected (capping to 10):");
-        for (i, err) in failed_outputs.iter().take(10).enumerate() {
-            eprintln!("Error {i}: {}", err.error);
-        }
-    }
+    log_failed_requests(outputs);

    // Compute peak concurrent requests from start_time + latency windows
    let successful_outputs: Vec<&RequestFuncOutput> =
@@ -6,6 +6,7 @@ use std::sync::Arc;
 use std::time::Instant;

 use indicatif::{ProgressBar, ProgressStyle};
+use thiserror_ext::AsReport as _;
 use tokio::sync::Semaphore;

 use crate::backends::{Backend, RequestFuncInput, RequestFuncOutput, get_backend};
@@ -72,9 +73,13 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
    let (model_id, model_name) = if let Some(ref m) = config.model {
        (m.clone(), config.model_name.clone())
    } else {
-        println!("Model not specified, fetching first model from server...");
+        tracing::info!(base_url = %config.base_url, "fetching first model from server");
        let (name, id) = get_first_model(&config.base_url, &client, &config.extra_headers).await?;
-        println!("First model name: {name}, first model id: {id}");
+        tracing::info!(
+            model_name = name,
+            model_id = id,
+            "selected first model from server"
+        );
        (id, Some(name))
    };

@@ -83,15 +88,19 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
        None
    } else {
        let tid = config.tokenizer_id.as_deref().unwrap_or(&model_id);
-        println!("Loading tokenizer: {tid}");
+        tracing::info!(tokenizer = tid, "loading tokenizer");
        let server_info = Some((config.base_url.as_str(), model_id.as_str()));
-        let t = crate::tokenizer::load_tokenizer(tid, config.trust_remote_code, server_info)?;
-        println!("Tokenizer loaded successfully.");
+        let t =
+            crate::tokenizer::load_tokenizer(tid, config.trust_remote_code, server_info).await?;
        Some(t)
    };

    // Generate/load conversations
-    println!("Generating multi-turn conversations...");
+    tracing::info!(
+        dataset = ?config.dataset_name,
+        conversations = config.num_prompts,
+        "generating multi-turn conversations"
+    );
    let gen_start = Instant::now();

    let mut conversations = match config.dataset_name {
@@ -131,7 +140,7 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
            let path = match config.dataset_path.as_deref() {
                Some(p) => p,
                None => {
-                    downloaded = crate::datasets::sharegpt::download_sharegpt_dataset()?;
+                    downloaded = crate::datasets::sharegpt::download_sharegpt_dataset().await?;
                    downloaded.as_str()
                }
            };
@@ -179,8 +188,11 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
        let (filtered_conversations, filtered_turns) =
            filter_turns_by_max_model_len(&mut conversations, max_model_len, no_history);
        if filtered_turns > 0 || filtered_conversations > 0 {
-            println!(
-                "Filtered {filtered_turns} turn(s) and {filtered_conversations} conversation(s) above --max-model-len {max_model_len}."
+            tracing::info!(
+                filtered_turns,
+                filtered_conversations,
+                max_model_len,
+                "filtered conversations above maximum model length"
            );
        }
        if conversations.is_empty() {
@@ -192,11 +204,11 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json

    let gen_elapsed = gen_start.elapsed();
    let total_turns: usize = conversations.iter().map(|c| c.turns.len()).sum();
-    println!(
-        "Generated {} conversations ({} total turns) in {:.2}s",
-        conversations.len(),
+    tracing::info!(
+        conversations = conversations.len(),
        total_turns,
-        gen_elapsed.as_secs_f64()
+        elapsed_seconds = gen_elapsed.as_secs_f64(),
+        "generated multi-turn conversations"
    );

    // Log prefix sharing info
@@ -208,18 +220,15 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
        let conv_tokens =
            (real_input_len as f64 * config.multi_turn_prefix_conversation_ratio).floor() as usize;
        let unique_tokens = real_input_len.saturating_sub(global_tokens + conv_tokens);
-        println!(
-            "User message prefix sharing: {:.0}% global ({} tokens), {:.0}% per-conversation ({} tokens), {:.0}% unique ({} tokens)",
-            config.multi_turn_prefix_global_ratio * 100.0,
+        tracing::info!(
+            global_ratio = config.multi_turn_prefix_global_ratio,
            global_tokens,
-            config.multi_turn_prefix_conversation_ratio * 100.0,
-            conv_tokens,
-            (1.0 - config.multi_turn_prefix_global_ratio
-                - config.multi_turn_prefix_conversation_ratio)
-                * 100.0,
+            conversation_ratio = config.multi_turn_prefix_conversation_ratio,
+            conversation_tokens = conv_tokens,
            unique_tokens,
+            history_accumulation = false,
+            "configured multi-turn prefix sharing"
        );
-        println!("No history accumulation: each turn sends fixed-length prompt only.");
    }

    if config.dry_run {
@@ -253,7 +262,7 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
            ..Default::default()
        };

-        println!("Starting initial single prompt test run...");
+        tracing::info!("starting initial single-prompt test run");
        let test_output = crate::ready_checker::wait_for_endpoint(
            config.backend,
            &client,
@@ -268,7 +277,7 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
                test_output.error
            )));
        }
-        println!("Initial test run completed.");
+        tracing::info!("initial single-prompt test run completed");
    }

    // For random datasets in multi-turn mode, auto-set min_tokens to enforce
@@ -283,9 +292,10 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
                "min_tokens".to_string(),
                serde_json::json!(config.random_output_len),
            );
-            println!(
-                "Auto-setting min_tokens={} for multi-turn random dataset (use --extra-body to override)",
-                config.random_output_len
+            tracing::info!(
+                min_tokens = config.random_output_len,
+                dataset = "random",
+                "set minimum output tokens for multi-turn dataset"
            );
        }
        Some(body)
@@ -297,7 +307,7 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
    let spec_decode_before =
        fetch_spec_decode_metrics(&config.base_url, &client, &config.extra_headers).await;
    if spec_decode_before.is_some() {
-        println!("Speculative decoding detected, will collect metrics.");
+        tracing::info!("detected speculative decoding; collecting metrics");
    }

    // Start profiler if requested (immediate mode — no batch threshold)
@@ -330,10 +340,13 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
    };

    // Main benchmark
-    println!("Starting multi-turn benchmark...");
-    println!("Conversations: {}", conversations.len());
-    println!("Concurrency: {concurrency}");
-    println!("Inter-turn delay: {} ms", config.multi_turn_delay_ms);
+    tracing::info!(
+        conversations = conversations.len(),
+        total_turns,
+        concurrency,
+        inter_turn_delay_ms = config.multi_turn_delay_ms,
+        "starting multi-turn benchmark"
+    );

    let max_turn_count = conversations.iter().map(|c| c.turns.len()).max().unwrap_or(0);

@@ -364,11 +377,12 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
    );
    if let Some(modules) = config.lora_modules.as_ref() {
        let names: Vec<&str> = modules.iter().map(|s| s.as_ref()).collect();
-        println!(
-            "LoRA adapters ({}): {:?} [assignment={:?}, scope=conversation]",
-            modules.len(),
-            names,
-            config.lora_assignment
+        tracing::info!(
+            adapters = modules.len(),
+            names = ?names,
+            assignment = ?config.lora_assignment,
+            scope = "conversation",
+            "assigned LoRA adapters"
        );
    }

@@ -433,7 +447,7 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
        match handle.await {
            Ok(output) => all_outputs.push(output),
            Err(e) => {
-                eprintln!("Conversation task panicked: {e}");
+                tracing::error!(error = %e.as_report(), "conversation task panicked");
            }
        }
    }
@@ -453,7 +467,7 @@ pub async fn run_multi_turn_benchmark(config: &BenchConfig) -> Result<serde_json
    if let Some((cancel_tx, task)) = profile_task {
        let _ = cancel_tx.send(());
        if let Err(e) = task.await {
-            eprintln!("WARNING: Profile background task failed: {e}");
+            tracing::error!(error = %e.as_report(), "profiler background task failed");
        }
    }

@@ -603,7 +603,7 @@ fn add_metric_stats(
 pub fn save_result(json: &Value, file_path: &str) -> Result<()> {
    let content = serde_json::to_string(json)?;
    std::fs::write(file_path, content)?;
-    println!("Results saved to {file_path}");
+    tracing::info!(path = file_path, "saved benchmark results");
    Ok(())
 }

@@ -618,7 +618,7 @@ pub fn append_result(json: &Value, file_path: &str) -> Result<()> {
        file.write_all(b"\n")?;
    }
    file.write_all(content.as_bytes())?;
-    println!("Results appended to {file_path}");
+    tracing::info!(path = file_path, "appended benchmark results");
    Ok(())
 }

@@ -23,7 +23,11 @@ pub async fn wait_for_endpoint(
    let backend = get_backend(backend)?;
    let deadline = Instant::now() + std::time::Duration::from_secs(timeout_seconds);

-    println!("Waiting for endpoint to become up in {timeout_seconds}s");
+    tracing::info!(
+        timeout_seconds,
+        retry_interval,
+        "waiting for endpoint readiness"
+    );

    let pb = ProgressBar::new(timeout_seconds);
    pb.set_style(
@@ -53,7 +57,9 @@ pub async fn wait_for_endpoint(
            Ok(output) => {
                let err = output.error.clone();
                let err_last_line = err.lines().last().unwrap_or(&err);
-                eprintln!("Endpoint is not ready. Error='{err_last_line}'");
+                pb.suspend(|| {
+                    tracing::warn!(error = err_last_line, "endpoint is not ready");
+                });
                last_error = err;
            }
            Err(e) => {
@@ -16,7 +16,7 @@ async fn reset_prefix_cache(base_url: &str) -> Result<()> {
        .await
        .map_err(|e| BenchError::Backend(format!("Failed to reset prefix cache: {e}")))?;
    if resp.status().is_success() {
-        println!("Prefix cache reset successfully.");
+        tracing::info!(url = %url, "reset prefix cache");
    } else {
        let status = resp.status();
        let body = resp.text().await.unwrap_or_default();
@@ -199,12 +199,17 @@ pub fn load_builtin_tiktoken(encoding: &str) -> Result<TiktokenTokenizer> {
        }
    };
    let bpe = bpe.map_err(|e| BenchError::Tokenizer(format!("Failed to load {encoding}: {e}")))?;
-    println!("Tokenizer: Built-in tiktoken {encoding} (vocab_size={vocab_size})");
+    tracing::info!(
+        encoding,
+        kind = "built-in-tiktoken",
+        vocab_size,
+        "loaded tokenizer"
+    );
    Ok(TiktokenTokenizer::from_builtin_bpe(bpe, vocab_size))
 }

 /// Try to load a tiktoken tokenizer from a local directory or HuggingFace model repo.
-pub fn try_load_tiktoken(model_id: &str) -> Result<TiktokenTokenizer> {
+pub async fn try_load_tiktoken(model_id: &str) -> Result<TiktokenTokenizer> {
    // Phase 1: If model_id is a local directory, look for tiktoken files there
    let local_dir = Path::new(model_id);
    if local_dir.is_dir() {
@@ -212,7 +217,7 @@ pub fn try_load_tiktoken(model_id: &str) -> Result<TiktokenTokenizer> {
    }

    // Phase 2: Fall back to HuggingFace Hub download
-    try_load_tiktoken_from_hf(model_id)
+    try_load_tiktoken_from_hf(model_id).await
 }

 /// Common tiktoken model filenames to search for.
@@ -247,25 +252,28 @@ fn try_load_tiktoken_from_dir(dir: &Path, model_id: &str) -> Result<TiktokenToke
 }

 /// Load a tiktoken tokenizer from a HuggingFace model repo.
-fn try_load_tiktoken_from_hf(model_id: &str) -> Result<TiktokenTokenizer> {
-    let repo = crate::hub::HubRepo::model(model_id.to_string());
+async fn try_load_tiktoken_from_hf(model_id: &str) -> Result<TiktokenTokenizer> {
+    let repo = crate::hub::HubRepo::model(model_id.to_string()).map_err(BenchError::Tokenizer)?;

-    let model_path = repo
-        .get("tiktoken.model")
-        .or_else(|_| repo.get("qwen.tiktoken"))
-        .or_else(|_| repo.get("vocab.tiktoken"))
-        .map_err(|_| {
-            BenchError::Tokenizer(format!("No tiktoken model file found for '{model_id}'"))
-        })?;
+    let mut model_path = None;
+    for filename in TIKTOKEN_MODEL_FILENAMES {
+        if let Ok(path) = repo.get(filename).await {
+            model_path = Some(path);
+            break;
+        }
+    }
+    let model_path = model_path.ok_or_else(|| {
+        BenchError::Tokenizer(format!("No tiktoken model file found for '{model_id}'"))
+    })?;

    let num_base_tokens = count_base_tokens(&model_path)?;

-    let config = match repo.get("tokenizer_config.json") {
+    let config = match repo.get("tokenizer_config.json").await {
        Ok(config_path) => read_tokenizer_config(&config_path),
        Err(_) => None,
    };

-    let pattern = extract_pat_str_from_repo(&repo);
+    let pattern = extract_pat_str_from_repo(&repo).await;

    build_tiktoken(model_id, &model_path, config, pattern, num_base_tokens)
 }
@@ -309,15 +317,16 @@ fn build_tiktoken(
        }
    }

-    println!(
-        "Loading tiktoken model for '{model_id}' (base={}, special={}, pat={})...",
-        num_base_tokens,
-        all_special_tokens.len(),
-        if pattern.is_some() {
+    tracing::info!(
+        model = model_id,
+        base_tokens = num_base_tokens,
+        special_tokens = all_special_tokens.len(),
+        pattern = if pattern.is_some() {
            "custom"
        } else {
            "default"
        },
+        "loading tiktoken model"
    );

    TiktokenTokenizer::from_file(
@@ -397,9 +406,12 @@ fn extract_pat_str_from_local_dir(dir: &Path) -> Option<String> {

 /// Try to download the Python tokenizer source file and extract pat_str via regex.
 /// Returns None if unavailable or unparsable.
-fn extract_pat_str_from_repo(repo: &crate::hub::HubRepo) -> Option<String> {
+async fn extract_pat_str_from_repo(repo: &crate::hub::HubRepo) -> Option<String> {
    // Try common Python tokenizer filenames
-    let py_path = repo.get("tokenization_kimi.py").or_else(|_| repo.get("tokenizer.py")).ok()?;
+    let py_path = match repo.get("tokenization_kimi.py").await {
+        Ok(path) => path,
+        Err(_) => repo.get("tokenizer.py").await.ok()?,
+    };

    let source = std::fs::read_to_string(&py_path).ok()?;

@@ -438,9 +450,9 @@ fn extract_pat_str_from_source(source: &str) -> Option<String> {

                if !fragments.is_empty() {
                    let pattern = fragments.join("|");
-                    println!(
-                        "Extracted pat_str from Python source: {} fragments",
-                        fragments.len()
+                    tracing::debug!(
+                        fragments = fragments.len(),
+                        "extracted tiktoken pattern from Python source"
                    );
                    return Some(pattern);
                }
@@ -2,8 +2,10 @@
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 use std::collections::HashSet;
+use std::future::Future;
 use std::path::Path;

+use thiserror_ext::AsReport as _;
 use tokenizers::Tokenizer;

 use crate::error::{BenchError, Result};
@@ -18,7 +20,8 @@ pub enum TokenizerKind {

 /// Server-side tokenizer using vLLM's /tokenize and /detokenize endpoints.
 pub struct ServerTokenizer {
-    client: reqwest::blocking::Client,
+    client: reqwest::Client,
+    runtime: tokio::runtime::Handle,
    tokenize_url: String,
    detokenize_url: String,
    model: String,
@@ -27,8 +30,8 @@ pub struct ServerTokenizer {

 impl ServerTokenizer {
    /// Create a new server tokenizer and verify connectivity.
-    pub fn new(base_url: &str, model: &str) -> Result<Self> {
-        let client = reqwest::blocking::Client::builder()
+    pub async fn new(base_url: &str, model: &str) -> Result<Self> {
+        let client = reqwest::Client::builder()
            .timeout(std::time::Duration::from_secs(30))
            .build()
            .map_err(|e| BenchError::Tokenizer(format!("Failed to build HTTP client: {e}")))?;
@@ -38,6 +41,7 @@ impl ServerTokenizer {

        let st = Self {
            client,
+            runtime: tokio::runtime::Handle::current(),
            tokenize_url,
            detokenize_url,
            model: model.to_string(),
@@ -45,7 +49,7 @@ impl ServerTokenizer {
        };

        // Probe the endpoint to verify it works and discover vocab size
-        let test_tokens = st.encode_inner("test")?;
+        let test_tokens = st.encode_async("test").await?;
        let max_id = test_tokens.iter().copied().max().unwrap_or(0);
        let estimated_vocab = (max_id * 2).max(131072);

@@ -56,6 +60,10 @@ impl ServerTokenizer {
    }

    fn encode_inner(&self, text: &str) -> Result<Vec<u32>> {
+        self.block_on(self.encode_async(text))
+    }
+
+    async fn encode_async(&self, text: &str) -> Result<Vec<u32>> {
        let payload = serde_json::json!({
            "model": self.model,
            "prompt": text,
@@ -66,6 +74,7 @@ impl ServerTokenizer {
            .post(&self.tokenize_url)
            .json(&payload)
            .send()
+            .await
            .map_err(|e| BenchError::Tokenizer(format!("Server tokenize failed: {e}")))?;

        if !resp.status().is_success() {
@@ -75,7 +84,7 @@ impl ServerTokenizer {
            )));
        }

-        let data: serde_json::Value = resp.json().map_err(|e| {
+        let data: serde_json::Value = resp.json().await.map_err(|e| {
            BenchError::Tokenizer(format!("Failed to parse tokenize response: {e}"))
        })?;

@@ -95,6 +104,10 @@ impl ServerTokenizer {
    }

    fn decode_inner(&self, ids: &[u32]) -> Result<String> {
+        self.block_on(self.decode_async(ids))
+    }
+
+    async fn decode_async(&self, ids: &[u32]) -> Result<String> {
        let payload = serde_json::json!({
            "model": self.model,
            "tokens": ids,
@@ -105,6 +118,7 @@ impl ServerTokenizer {
            .post(&self.detokenize_url)
            .json(&payload)
            .send()
+            .await
            .map_err(|e| BenchError::Tokenizer(format!("Server detokenize failed: {e}")))?;

        if !resp.status().is_success() {
@@ -114,7 +128,7 @@ impl ServerTokenizer {
            )));
        }

-        let data: serde_json::Value = resp.json().map_err(|e| {
+        let data: serde_json::Value = resp.json().await.map_err(|e| {
            BenchError::Tokenizer(format!("Failed to parse detokenize response: {e}"))
        })?;

@@ -123,6 +137,26 @@ impl ServerTokenizer {
            .map(|s| s.to_string())
            .ok_or_else(|| BenchError::Tokenizer("Missing 'prompt' in detokenize response".into()))
    }
+
+    fn block_on<T>(&self, future: impl Future<Output = Result<T>>) -> Result<T> {
+        if matches!(
+            self.runtime.runtime_flavor(),
+            tokio::runtime::RuntimeFlavor::CurrentThread
+        ) {
+            return Err(BenchError::Tokenizer(
+                "Server tokenizer fallback requires a multi-thread Tokio runtime".into(),
+            ));
+        }
+
+        // Sync tokenizer calls can come from a Tokio worker or a Rayon worker.
+        // Tokio workers must enter a blocking region before re-entering the runtime;
+        // Rayon workers can drive the future directly with the saved runtime handle.
+        if tokio::runtime::Handle::try_current().is_ok() {
+            tokio::task::block_in_place(|| self.runtime.block_on(future))
+        } else {
+            self.runtime.block_on(future)
+        }
+    }
 }

 // --- TokenizerKind methods ---
@@ -192,7 +226,7 @@ impl TokenizerKind {
 /// 3. Server-side /tokenize + /detokenize endpoints
 ///
 /// `server_info` is `Some((base_url, model))` to enable server-side fallback.
-pub fn load_tokenizer(
+pub async fn load_tokenizer(
    model_id: &str,
    _trust_remote_code: bool,
    server_info: Option<(&str, &str)>,
@@ -212,31 +246,48 @@ pub fn load_tokenizer(
    }

    // 1. Try local HuggingFace tokenizer (tokenizer.json)
-    match try_load_local(model_id) {
+    match try_load_local(model_id).await {
        Ok(tok) => {
-            println!("Tokenizer: Local (vocab_size={})", tok.get_vocab_size(true));
+            tracing::info!(
+                model = model_id,
+                kind = "local",
+                vocab_size = tok.get_vocab_size(true),
+                "loaded tokenizer"
+            );
            Ok(TokenizerKind::Local(Box::new(tok)))
        }
        Err(local_err) => {
            // 2. Try tiktoken format
-            println!("No tokenizer.json for '{model_id}', trying tiktoken format...");
-            match crate::tiktoken::try_load_tiktoken(model_id) {
+            tracing::info!(
+                model = model_id,
+                error = %local_err.as_report(),
+                "local tokenizer unavailable; trying tiktoken"
+            );
+            match crate::tiktoken::try_load_tiktoken(model_id).await {
                Ok(tok) => {
-                    println!("Tokenizer: Tiktoken (vocab_size={})", tok.vocab_size());
+                    tracing::info!(
+                        model = model_id,
+                        kind = "tiktoken",
+                        vocab_size = tok.vocab_size(),
+                        "loaded tokenizer"
+                    );
                    Ok(TokenizerKind::Tiktoken(tok))
                }
                Err(tiktoken_err) => {
                    // 3. Try server-side fallback
                    if let Some((base_url, model)) = server_info {
-                        println!(
-                            "Tiktoken also not available ({tiktoken_err}), \
-                             trying server-side tokenization..."
+                        tracing::info!(
+                            model = model_id,
+                            error = %tiktoken_err.as_report(),
+                            "tiktoken unavailable; trying server-side tokenization"
                        );
-                        match ServerTokenizer::new(base_url, model) {
+                        match ServerTokenizer::new(base_url, model).await {
                            Ok(srv) => {
-                                println!(
-                                    "Tokenizer: Server (vocab_size≈{})",
-                                    srv.cached_vocab_size
+                                tracing::info!(
+                                    model = model_id,
+                                    kind = "server",
+                                    vocab_size = srv.cached_vocab_size,
+                                    "loaded tokenizer"
                                );
                                return Ok(TokenizerKind::Server(srv));
                            }
@@ -264,7 +315,7 @@ pub fn load_tokenizer(
 }

 /// Try loading tokenizer.json from local path or HuggingFace Hub.
-fn try_load_local(model_id: &str) -> Result<Tokenizer> {
+async fn try_load_local(model_id: &str) -> Result<Tokenizer> {
    // 1. Try local directory with tokenizer.json
    let local_path = Path::new(model_id).join("tokenizer.json");
    if local_path.exists() {
@@ -290,11 +341,37 @@ fn try_load_local(model_id: &str) -> Result<Tokenizer> {
    }

    // 4. Download from HuggingFace Hub (hf-hub handles auth via HF_TOKEN / cached token)
-    let repo = crate::hub::HubRepo::model(model_id.to_string());
+    let repo = crate::hub::HubRepo::model(model_id.to_string()).map_err(BenchError::Tokenizer)?;
    let tokenizer_path = repo
        .get("tokenizer.json")
+        .await
        .map_err(|e| BenchError::Tokenizer(format!("No tokenizer.json for '{model_id}': {e}")))?;

    Tokenizer::from_file(&tokenizer_path)
        .map_err(|e| BenchError::Tokenizer(format!("Failed to load downloaded tokenizer: {e}")))
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    async fn test_server_tokenizer_sync_bridge() {
+        let tokenizer = std::sync::Arc::new(ServerTokenizer {
+            client: reqwest::Client::new(),
+            runtime: tokio::runtime::Handle::current(),
+            tokenize_url: String::new(),
+            detokenize_url: String::new(),
+            model: String::new(),
+            cached_vocab_size: 0,
+        });
+
+        assert_eq!(tokenizer.block_on(async { Ok(1) }).unwrap(), 1);
+
+        let (tx, rx) = tokio::sync::oneshot::channel();
+        rayon::spawn(move || {
+            let _ = tx.send(tokenizer.block_on(async { Ok(2) }));
+        });
+        assert_eq!(rx.await.unwrap().unwrap(), 2);
+    }
+}
@@ -74,7 +74,7 @@ impl DefaultChatOutputProcessor {
            Box::new(CombinedParser::new(reasoning_parser, tool_parser)) as Box<dyn UnifiedParser>
        };

-        apply_structural_tag_constraint(request, parser.structural_tag_model())?;
+        apply_structural_tag_constraint(request, parser.structural_tag_builder())?;

        if parser.preserve_special_tokens() {
            request.decode_options.skip_special_tokens = false;
@@ -7,7 +7,8 @@ use thiserror_ext::AsReport;
 use vllm_engine_core_client::protocol::structured_outputs::{
    StructuredOutputBackend, StructuredOutputsParams,
 };
-use vllm_parser::tool::StructuralTagModel;
+use vllm_parser::tool::StructuralTagBuilder;
+use xgrammar_structural_tag::builders::StructuralTagOptions;
 use xgrammar_structural_tag::{
    FunctionDefinition, FunctionToolParam, ToolChoice as StructuralTagToolChoice, ToolParam,
    build_structural_tag,
@@ -20,9 +21,9 @@ use crate::{Error, Result as ChatResult};
 /// support and the request's tool choice.
 pub(super) fn apply_structural_tag_constraint(
    request: &mut ChatRequest,
-    model: Option<StructuralTagModel>,
+    builder: Option<&dyn StructuralTagBuilder>,
 ) -> ChatResult<()> {
-    let Some(model) = model else {
+    let Some(builder) = builder else {
        return Ok(());
    };
    let Some(tool_choice) = structural_tag_tool_choice(request) else {
@@ -42,11 +43,16 @@ pub(super) fn apply_structural_tag_constraint(
        })
        .collect::<Vec<_>>();

-    let structural_tag = build_structural_tag(model, &tools, tool_choice, false)
-        .and_then(|tag| tag.to_json_string())
-        .map_err(|error| Error::StructuralTag {
-            message: error.to_report_string(),
-        })?;
+    let structural_tag = build_structural_tag(
+        builder,
+        &tools,
+        tool_choice,
+        StructuralTagOptions::default().with_reasoning(false),
+    )
+    .and_then(|tag| tag.to_json_string())
+    .map_err(|error| Error::StructuralTag {
+        message: error.to_report_string(),
+    })?;

    // Overwrite any existing structured output settings with the structural tag constraint.
    request.sampling_params.structured_outputs = Some(StructuredOutputsParams {
@@ -141,7 +147,7 @@ mod tests {
        let mut request = request(ChatToolChoice::Auto, vec![chat_tool("search", Some(true))]);
        let parser = qwen3_coder_parser(&request.tools);

-        apply_structural_tag_constraint(&mut request, parser.structural_tag_model())
+        apply_structural_tag_constraint(&mut request, parser.structural_tag_builder())
            .expect("structural tag should build");

        let tag = structural_tag_value(&request);
@@ -154,7 +160,7 @@ mod tests {
        let mut request = request(ChatToolChoice::Auto, vec![chat_tool("search", None)]);
        let parser = qwen3_coder_parser(&request.tools);

-        apply_structural_tag_constraint(&mut request, parser.structural_tag_model())
+        apply_structural_tag_constraint(&mut request, parser.structural_tag_builder())
            .expect("structural tag decision should succeed");

        assert!(request.sampling_params.structured_outputs.is_none());
@@ -169,7 +175,7 @@ mod tests {
        });
        let parser = qwen3_coder_parser(&request.tools);

-        apply_structural_tag_constraint(&mut request, parser.structural_tag_model())
+        apply_structural_tag_constraint(&mut request, parser.structural_tag_builder())
            .expect("structural tag should build");

        let params = structured_outputs(&request);
@@ -184,7 +190,7 @@ mod tests {
        let mut request = request(ChatToolChoice::Required, vec![chat_tool("search", None)]);
        let parser = qwen3_coder_parser(&request.tools);

-        apply_structural_tag_constraint(&mut request, parser.structural_tag_model())
+        apply_structural_tag_constraint(&mut request, parser.structural_tag_builder())
            .expect("structural tag should build");

        let tag = structural_tag_value(&request);
@@ -201,7 +207,7 @@ mod tests {
        });
        let parser = qwen3_coder_parser(&request.tools);

-        apply_structural_tag_constraint(&mut request, parser.structural_tag_model())
+        apply_structural_tag_constraint(&mut request, parser.structural_tag_builder())
            .expect("structural tag should build");

        let params = structured_outputs(&request);
@@ -221,7 +227,7 @@ mod tests {
        );
        let parser = qwen3_coder_parser(&request.tools);

-        apply_structural_tag_constraint(&mut request, parser.structural_tag_model())
+        apply_structural_tag_constraint(&mut request, parser.structural_tag_builder())
            .expect("structural tag should build");

        let tag = structural_tag_value(&request).to_string();
@@ -234,7 +240,7 @@ mod tests {
        let mut request = request(ChatToolChoice::None, vec![chat_tool("search", Some(true))]);
        let parser = qwen3_coder_parser(&request.tools);

-        apply_structural_tag_constraint(&mut request, parser.structural_tag_model())
+        apply_structural_tag_constraint(&mut request, parser.structural_tag_builder())
            .expect("structural tag decision should succeed");

        assert!(request.sampling_params.structured_outputs.is_none());
@@ -249,7 +255,7 @@ mod tests {
        });
        let parser = qwen3_coder_parser(&request.tools);

-        apply_structural_tag_constraint(&mut request, parser.structural_tag_model())
+        apply_structural_tag_constraint(&mut request, parser.structural_tag_builder())
            .expect("structural tag decision should succeed");

        let params = structured_outputs(&request);
@@ -26,18 +26,21 @@ const RESET: &str = "\x1b[0m";
 const VLLM_TIME_FORMAT: &[time::format_description::FormatItem<'static>] =
    format_description!("[month]-[day] [hour]:[minute]:[second]");

-const PROCESS_LABEL: &str = "RustFrontend";
-
 /// Install the process-wide vLLM-style tracing subscriber for the CLI binary.
-pub(crate) fn init_tracing() {
+pub(crate) fn init_tracing(process_label: &str) {
    let filter = build_targets_filter(
        env::var("VLLM_LOGGING_LEVEL").ok().as_deref(),
        env::var("RUST_LOG").ok().as_deref(),
    );
-    let formatter = VllmEventFormatter::new();
+    let formatter = VllmEventFormatter::new(process_label);

    let _ = tracing_subscriber::registry()
-        .with(tracing_subscriber::fmt::layer().event_format(formatter).with_filter(filter))
+        .with(
+            tracing_subscriber::fmt::layer()
+                .event_format(formatter)
+                .with_writer(std::io::stderr)
+                .with_filter(filter),
+        )
        .try_init();
 }

@@ -94,9 +97,9 @@ struct VllmEventFormatter {
 }

 impl VllmEventFormatter {
-    fn new() -> Self {
+    fn new(process_label: &str) -> Self {
        Self {
-            prefix: format!("({} pid={})", PROCESS_LABEL, process::id()),
+            prefix: format!("({process_label} pid={})", process::id()),
            timer: VllmLocalTimer::default(),
        }
    }
@@ -291,6 +294,13 @@ fn map_python_log_level(level: &str) -> LevelFilter {
 mod tests {
    use super::*;

+    #[test]
+    fn formatter_prefix_uses_process_label() {
+        let formatter = VllmEventFormatter::new("Bench");
+
+        assert_eq!(formatter.prefix, format!("(Bench pid={})", process::id()));
+    }
+
    #[test]
    fn rust_log_target_overrides_are_merged_with_vllm_default_level() {
        let filter = build_targets_filter(Some("DEBUG"), Some("hyper=warn,tower=error"));
@@ -5,6 +5,7 @@ mod cli;
 mod logging;

 use std::env;
+use std::ffi::OsStr;
 use std::process::ExitStatus;

 use anyhow::{Context, Result, anyhow, bail};
@@ -82,7 +83,14 @@ fn shutdown_signal() -> CancellationToken {
 }

 fn main() -> Result<()> {
-    logging::init_tracing();
+    let process_label =
+        match env::args_os().nth(1).as_deref().and_then(OsStr::to_str).unwrap_or_default() {
+            "bench" => "Bench",
+            "serve" | "frontend" => "RustFrontend",
+            _ => "Rust",
+        };
+    logging::init_tracing(process_label);
+
    let cli = Cli::parse();

    let mut runtime = tokio::runtime::Builder::new_multi_thread();
@@ -460,6 +460,12 @@ impl EngineCoreClient {
        self.inner.is_healthy()
    }

+    /// Subscribe to engine health changes. The current value is `true` while
+    /// the client is healthy and changes permanently to `false` on failure.
+    pub fn subscribe_health(&self) -> tokio::sync::watch::Receiver<bool> {
+        self.inner.subscribe_health()
+    }
+
    /// Return the first persistent health error observed by the client, if any.
    pub fn health_error(&self) -> Option<Arc<Error>> {
        self.inner.health_error()
@@ -9,7 +9,7 @@ use arc_swap::ArcSwapOption;
 use parking_lot::Mutex;
 use thiserror_ext::AsReport as _;
 use tokio::runtime::Handle;
-use tokio::sync::mpsc;
+use tokio::sync::{mpsc, watch};
 use tracing::{debug, info, trace, warn};
 use vllm_metrics::METRICS;
 use zeromq::RouterSendHalf;
@@ -36,6 +36,7 @@ pub(crate) struct ClientInner {
    request_reg: Mutex<RequestRegistry>,
    utility_reg: Mutex<UtilityRegistry>,
    health_error: ArcSwapOption<Error>,
+    health_tx: watch::Sender<bool>,
 }

 impl ClientInner {
@@ -57,6 +58,7 @@ impl ClientInner {
            request_reg: Mutex::new(RequestRegistry::new(engines)),
            utility_reg: Mutex::new(UtilityRegistry::default()),
            health_error: ArcSwapOption::empty(),
+            health_tx: watch::Sender::new(true),
        }
    }

@@ -169,6 +171,7 @@ impl ClientInner {
    /// persistent health error.
    pub fn close_registries(&self, error: Arc<Error>) {
        let persistent_error = self.record_health_error(error);
+        self.publish_unhealthy();
        let request_senders = self.request_reg.lock().close();
        let utility_senders = self.utility_reg.lock().close();

@@ -191,6 +194,12 @@ impl ClientInner {
        self.health_error.load().is_none()
    }

+    /// Subscribe to engine health changes. The current value is `true` while
+    /// the client is healthy and changes permanently to `false` on failure.
+    pub fn subscribe_health(&self) -> watch::Receiver<bool> {
+        self.health_tx.subscribe()
+    }
+
    /// Resolve one utility output to the waiting caller. Returns `true` if a
    /// waiting caller existed.
    pub fn resolve_utility_output(&self, output: UtilityOutput) -> bool {
@@ -280,6 +289,11 @@ impl ClientInner {
            .expect("health error must be recorded before registries close")
    }

+    /// Publish the sticky healthy-to-unhealthy transition.
+    fn publish_unhealthy(&self) {
+        self.health_tx.send_if_modified(|healthy| std::mem::replace(healthy, false));
+    }
+
    /// Assert there is a recorded health error and return a `Shared` variant
    /// wrapping it for error returns when the client is already closed.
    fn closed_error(&self) -> Error {
@@ -461,13 +475,18 @@ mod tests {
    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn close_registries_records_first_health_error_only() {
        let inner = test_inner().await;
+        let mut health = inner.subscribe_health();
+        assert!(*health.borrow());

        inner.close_registries(Arc::new(Error::EngineCoreDead));
+        health.changed().await.expect("health sender remains open");
        assert!(!inner.is_healthy());
+        assert!(!*health.borrow());
        assert!(matches!(
            inner.health_error().as_deref(),
            Some(Error::EngineCoreDead)
        ));
+        assert!(!*inner.subscribe_health().borrow());

        inner.close_registries(Arc::new(client_closed!("shutdown")));
        assert!(matches!(
@@ -269,6 +269,19 @@ impl WireLogprobs {
            );
        }

+        // Empty position lists may be encoded as either [0, 0] or [0, k + 1].
+        if token_ids.rows == 0 {
+            return Ok(Logprobs {
+                positions: Vec::new(),
+            });
+        }
+        if token_ids.cols == 0 {
+            bail_ext_value_decode!(
+                "{field_prefix}: zero-column logprobs payload with {} rows",
+                token_ids.rows
+            );
+        }
+
        let mut positions = Vec::with_capacity(token_ids.rows);
        for ((token_ids_row, logprobs_row), sampled_rank) in token_ids
            .data
@@ -303,3 +303,49 @@ fn rejects_non_none_cu_num_generated_tokens() {
        "messagepack ext value decode failed: new_logprobs.cu_num_generated_tokens: expected None for per-request engine-core logprobs payload, got [0, 1]"
    );
 }
+
+#[test]
+fn decodes_zero_row_logprobs_as_empty() {
+    for shape in [[0usize, 0], [0, 3]] {
+        let frames = vec![Bytes::from(encode_value(&output_wire_with_custom_fields(
+            None,
+            Some(Value::Array(vec![
+                ndarray_value("<i8", &shape, Value::Ext(3, Vec::new())),
+                ndarray_value("<f4", &shape, Value::Ext(3, Vec::new())),
+                ndarray_value("<i8", &[0], Value::Ext(3, Vec::new())),
+                Value::Nil,
+            ])),
+        )))];
+        let decoded = decode_engine_core_outputs(&frames).unwrap().into_request_batch().unwrap();
+        let logprobs = decoded.outputs[0]
+            .new_prompt_logprobs_tensors
+            .clone()
+            .unwrap()
+            .into_direct()
+            .unwrap();
+        assert!(logprobs.is_empty());
+    }
+}
+
+#[test]
+fn rejects_zero_column_logprobs_with_rows() {
+    let ranks = Value::Ext(3, vec![1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0]);
+    let frames = vec![Bytes::from(encode_value(&output_wire_with_custom_fields(
+        Some(Value::Array(vec![
+            ndarray_value("<i8", &[2, 0], Value::Ext(3, Vec::new())),
+            ndarray_value("<f4", &[2, 0], Value::Ext(3, Vec::new())),
+            ndarray_value("<i8", &[2], ranks),
+            Value::Nil,
+        ])),
+        None,
+    )))];
+
+    let error = decode_engine_core_outputs(&frames).unwrap_err();
+    let crate::error::Error::ExtValueDecode { message } = &error else {
+        panic!("expected ExtValueDecode");
+    };
+    assert_eq!(
+        message,
+        "new_logprobs: zero-column logprobs payload with 2 rows"
+    );
+}
@@ -4,7 +4,7 @@
 use std::sync::Arc;

 use vllm_parser::tool::{
-    Result, StructuralTagModel, Tool, ToolParser, ToolParserError, ToolParserOutput,
+    Result, StructuralTagBuilder, Tool, ToolParser, ToolParserError, ToolParserOutput,
 };
 use vllm_parser::unified::{
    UnifiedParser, UnifiedParserError, UnifiedParserEvent, UnifiedParserOutput,
@@ -85,8 +85,8 @@ impl<T: UnifiedParser> ToolParser for UnifiedToolParserAdapter<T> {
        self.inner.preserve_special_tokens()
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        self.inner.structural_tag_model()
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        self.inner.structural_tag_builder()
    }

    fn tool_call_id(&self, tool_index: usize) -> Option<&str> {
@@ -2,7 +2,7 @@
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 use super::{DeepSeekDsmlToolParser, DsmlTokens};
-use crate::tool::{Result, StructuralTagModel, Tool, ToolParser, ToolParserOutput};
+use crate::tool::{Result, StructuralTagBuilder, Tool, ToolParser, ToolParserOutput};

 /// Tool parser for DeepSeek V3.2 models.
 ///
@@ -47,8 +47,8 @@ impl ToolParser for DeepSeekV32ToolParser {
        true
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::DeepSeekV32)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::DeepSeekV32.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -2,7 +2,7 @@
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 use super::{DeepSeekDsmlToolParser, DsmlTokens};
-use crate::tool::{Result, StructuralTagModel, Tool, ToolParser, ToolParserOutput};
+use crate::tool::{Result, StructuralTagBuilder, Tool, ToolParser, ToolParserOutput};

 /// Tool parser for DeepSeek V4 models.
 ///
@@ -50,8 +50,8 @@ impl ToolParser for DeepSeekV4ToolParser {
        true
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::DeepSeekV4)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::DeepSeekV4.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -73,7 +73,7 @@ mod tests {

    use super::DeepSeekV4ToolParser;
    use crate::tool::test_utils::{collect_stream, test_tools};
-    use crate::tool::{StructuralTagModel, ToolParser, ToolParserTestExt as _};
+    use crate::tool::{ToolParser, ToolParserTestExt as _};

    fn build_tool_call(function_name: &str, params: &[(&str, &str)]) -> String {
        let params = params
@@ -91,13 +91,10 @@ mod tests {
    }

    #[test]
-    fn deepseek_v4_exposes_structural_tag_model() {
+    fn deepseek_v4_exposes_structural_tag_builder() {
        let parser = DeepSeekV4ToolParser::new(&test_tools());

-        assert_eq!(
-            parser.structural_tag_model(),
-            Some(StructuralTagModel::DeepSeekV4)
-        );
+        assert!(parser.structural_tag_builder().is_some());
    }

    #[test]
@@ -2,7 +2,7 @@
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 use super::{DeepSeekJsonFormat, DeepSeekJsonToolParser};
-use crate::tool::{Result, StructuralTagModel, Tool, ToolParser, ToolParserOutput};
+use crate::tool::{Result, StructuralTagBuilder, Tool, ToolParser, ToolParserOutput};

 /// Tool parser for DeepSeek V3 JSON-fenced tool calls.
 ///
@@ -35,8 +35,8 @@ impl ToolParser for DeepSeekV3ToolParser {
        Ok(Box::new(Self::new(tools)))
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::DeepSeekR1)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::DeepSeekR1.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -2,7 +2,7 @@
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 use super::{DeepSeekJsonFormat, DeepSeekJsonToolParser};
-use crate::tool::{Result, StructuralTagModel, Tool, ToolParser, ToolParserOutput};
+use crate::tool::{Result, StructuralTagBuilder, Tool, ToolParser, ToolParserOutput};

 /// Tool parser for DeepSeek V3.1 raw JSON tool calls.
 ///
@@ -31,8 +31,8 @@ impl ToolParser for DeepSeekV31ToolParser {
        Ok(Box::new(Self::new(tools)))
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::DeepSeekV31)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::DeepSeekV31.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -2,7 +2,7 @@
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 use super::{GlmXmlToolParser, Separator};
-use crate::tool::{Result, StructuralTagModel, Tool, ToolParser, ToolParserOutput};
+use crate::tool::{Result, StructuralTagBuilder, Tool, ToolParser, ToolParserOutput};

 /// Tool parser for GLM-4.7 MoE XML-style tool calls.
 ///
@@ -25,8 +25,8 @@ impl ToolParser for Glm47MoeToolParser {
        Ok(Box::new(Self::new(tools)))
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::Glm47)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::Glm47.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -10,7 +10,7 @@ use winnow::token::{literal, rest, take_until};
 use super::parameters::ToolSchemas;
 use super::utils::{MarkerScanState, parse_buffered_event, safe_text_len, take_until_marker};
 use super::{Result, ToolCallDelta, ToolParser, ToolParserOutput};
-use crate::tool::{StructuralTagModel, Tool};
+use crate::tool::{StructuralTagBuilder, Tool};

 const TOOL_CALLS_START: &str = "<tool_calls>";
 const TOOL_CALLS_END: &str = "</tool_calls>";
@@ -116,8 +116,8 @@ impl ToolParser for HyV3ToolParser {
        Ok(Box::new(Self::new(tools)))
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::HyV3)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::HyV3.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -2,7 +2,7 @@
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 use super::{JsonToolCallConfig, JsonToolCallParser, JsonToolCallWhitespace};
-use crate::tool::{Result, StructuralTagModel, Tool, ToolParser, ToolParserOutput};
+use crate::tool::{Result, StructuralTagBuilder, Tool, ToolParser, ToolParserOutput};

 const HERMES_CONFIG: JsonToolCallConfig = JsonToolCallConfig {
    parser_name: "Hermes",
@@ -48,8 +48,8 @@ impl ToolParser for HermesToolParser {
        Ok(Box::new(Self::new(tools)))
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::Hermes)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::Hermes.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -12,7 +12,9 @@ use super::{
    argument_delta_event, tool_call_header_event,
 };
 use crate::tool::utils::{JsonObjectScanState, parse_buffered_event};
-use crate::tool::{Result, StructuralTagModel, Tool, ToolCallDelta, ToolParser, ToolParserOutput};
+use crate::tool::{
+    Result, StructuralTagBuilder, Tool, ToolCallDelta, ToolParser, ToolParserOutput,
+};

 #[derive(Debug, Clone, PartialEq, Eq)]
 enum LlamaJsonMode {
@@ -136,8 +138,8 @@ impl ToolParser for Llama3JsonToolParser {
        Ok(Box::new(Self::new(tools)))
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::Llama)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::Llama.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -2,7 +2,7 @@
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 use super::{JsonToolCallConfig, JsonToolCallParser, JsonToolCallWhitespace};
-use crate::tool::{Result, StructuralTagModel, Tool, ToolParser, ToolParserOutput};
+use crate::tool::{Result, StructuralTagBuilder, Tool, ToolParser, ToolParserOutput};

 const QWEN_XML_CONFIG: JsonToolCallConfig = JsonToolCallConfig {
    parser_name: "Qwen XML",
@@ -50,8 +50,8 @@ impl ToolParser for Qwen3XmlToolParser {
        Ok(Box::new(Self::new(tools)))
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::Qwen3)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::Qwen3.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -11,7 +11,7 @@ use winnow::token::{literal, rest, take_until, take_while};

 use super::utils::{JsonObjectScanState, parse_buffered_event, safe_text_len, take_json_object};
 use super::{Result, ToolCallDelta, ToolParser, ToolParserOutput};
-use crate::tool::{StructuralTagModel, Tool};
+use crate::tool::{StructuralTagBuilder, Tool};

 const TOOL_CALLS_START: &str = "<|tool_calls_section_begin|>";
 const TOOL_CALLS_END: &str = "<|tool_calls_section_end|>";
@@ -150,8 +150,8 @@ impl ToolParser for KimiK2ToolParser {
        true
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::Kimi)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::Kimi.builder())
    }

    fn tool_call_id(&self, tool_index: usize) -> Option<&str> {
@@ -10,7 +10,7 @@ use winnow::token::{literal, rest, take_until};
 use super::parameters::ToolSchemas;
 use super::utils::{MarkerScanState, parse_buffered_event, safe_text_len, take_until_marker};
 use super::{Result, ToolCallDelta, ToolParser, ToolParserOutput};
-use crate::tool::{StructuralTagModel, Tool};
+use crate::tool::{StructuralTagBuilder, Tool};

 const TOOL_CALL_START: &str = "<minimax:tool_call>";
 const TOOL_CALL_END: &str = "</minimax:tool_call>";
@@ -115,8 +115,8 @@ impl ToolParser for MinimaxM2ToolParser {
        Ok(Box::new(Self::new(tools)))
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::Minimax)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::Minimax.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -36,7 +36,7 @@ pub use qwen_coder::Qwen3CoderToolParser;
 pub use seed_oss::SeedOssToolParser;
 use serde::{Deserialize, Serialize};
 use serde_json::Value;
-pub use xgrammar_structural_tag::Model as StructuralTagModel;
+pub use xgrammar_structural_tag::builders::StructuralTagBuilder;

 use crate::utils;

@@ -187,8 +187,8 @@ pub trait ToolParser: Send {
        false
    }

-    /// Return the xgrammar structural-tag model used for strict tool calling.
-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
+    /// Return the xgrammar structural-tag builder used for strict tool calling.
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
        None
    }

@@ -9,8 +9,8 @@ use winnow::token::{literal, take_until};

 use super::parameters::ToolSchemas;
 use super::utils::{MarkerScanState, parse_buffered_event, safe_text_len, take_until_marker};
-use super::{Result, StructuralTagModel, ToolCallDelta, ToolParser, ToolParserOutput};
-use crate::tool::Tool;
+use super::{Result, ToolCallDelta, ToolParser, ToolParserOutput};
+use crate::tool::{StructuralTagBuilder, Tool};

 const TOOL_CALL_START: &str = "<tool_call>";
 const TOOL_CALL_END: &str = "</tool_call>";
@@ -146,8 +146,8 @@ impl ToolParser for Qwen3CoderToolParser {
        Ok(Box::new(Self::new(tools)))
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        Some(StructuralTagModel::Qwen3Coder)
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        Some(xgrammar_structural_tag::Model::Qwen3Coder.builder())
    }

    fn parse_into(&mut self, chunk: &str, output: &mut ToolParserOutput) -> Result<()> {
@@ -294,7 +294,7 @@ mod tests {
    use serde_json::{Value, json};
    use thiserror_ext::AsReport;

-    use super::{Qwen3CoderToolParser, StructuralTagModel, ToolParser};
+    use super::{Qwen3CoderToolParser, ToolParser};
    use crate::tool::test_utils::{collect_stream, split_by_chars, test_tools};
    use crate::tool::{ToolParserOutput, ToolParserTestExt as _};

@@ -308,13 +308,10 @@ mod tests {
    }

    #[test]
-    fn qwen_coder_exposes_structural_tag_model() {
+    fn qwen_coder_exposes_structural_tag_builder() {
        let parser = Qwen3CoderToolParser::new(&test_tools());

-        assert_eq!(
-            parser.structural_tag_model(),
-            Some(StructuralTagModel::Qwen3Coder)
-        );
+        assert!(parser.structural_tag_builder().is_some());
    }

    #[test]
@@ -7,7 +7,7 @@ use vllm_tokenizer::DynTokenizer;

 use super::{Result, UnifiedParser, UnifiedParserError, UnifiedParserOutput};
 use crate::reasoning::ReasoningParser;
-use crate::tool::{StructuralTagModel, Tool, ToolParser, ToolParserOutput};
+use crate::tool::{StructuralTagBuilder, Tool, ToolParser, ToolParserOutput};

 /// Unified parser that composes existing reasoning and tool parsers.
 pub struct CombinedParser {
@@ -79,8 +79,8 @@ impl UnifiedParser for CombinedParser {
            || self.tool.as_ref().is_some_and(|parser| parser.preserve_special_tokens())
    }

-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
-        self.tool.as_ref().and_then(|parser| parser.structural_tag_model())
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
+        self.tool.as_ref().and_then(|parser| parser.structural_tag_builder())
    }

    fn tool_call_id(&self, tool_index: usize) -> Option<&str> {
@@ -269,10 +269,7 @@ mod tests {
    fn combined_parser_emits_tool_calls_from_visible_content() {
        let tool = Qwen3XmlToolParser::create(&test_tools()).unwrap();
        let mut parser = CombinedParser::new(None, Some(tool));
-        assert!(matches!(
-            parser.structural_tag_model(),
-            Some(crate::tool::StructuralTagModel::Qwen3)
-        ));
+        assert!(parser.structural_tag_builder().is_some());

        let output = collect(
            &mut parser,
@@ -16,7 +16,7 @@ use vllm_tokenizer::DynTokenizer;

 use crate::reasoning::ReasoningError;
 use crate::tool::{
-    StructuralTagModel, Tool, ToolCallDelta, ToolParserError, ToolParserEvent, ToolParserOutput,
+    StructuralTagBuilder, Tool, ToolCallDelta, ToolParserError, ToolParserEvent, ToolParserOutput,
 };

 /// Result alias for unified parser operations.
@@ -171,8 +171,8 @@ pub trait UnifiedParser: Send {
        false
    }

-    /// Return the xgrammar structural-tag model used for strict tool calling.
-    fn structural_tag_model(&self) -> Option<StructuralTagModel> {
+    /// Return the xgrammar structural-tag builder used for strict tool calling.
+    fn structural_tag_builder(&self) -> Option<&dyn StructuralTagBuilder> {
        None
    }

@@ -35,6 +35,7 @@ tokio-openssl.workspace = true
 tokio-stream.workspace = true
 tokio-util.workspace = true
 tonic.workspace = true
+tonic-health.workspace = true
 tonic-prost.workspace = true
 tower.workspace = true
 tower-http.workspace = true
@@ -0,0 +1,70 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+use tokio::sync::watch;
+use tokio_util::sync::CancellationToken;
+use tonic::server::NamedService;
+use tonic_health::ServingStatus;
+use tonic_health::server::HealthReporter;
+use tracing::{info, warn};
+
+use super::GenerateGrpcService;
+
+pub(crate) async fn monitor_health(
+    mut health_reporter: HealthReporter,
+    mut engine_health: watch::Receiver<bool>,
+    shutdown: CancellationToken,
+) {
+    let generate_service = GenerateGrpcService::NAME;
+    let status = ServingStatus::NotServing;
+    let health_event_first = tokio::select! {
+        result = engine_health.wait_for(|healthy| !*healthy) => {
+            match result {
+                Ok(_) => warn!(
+                    generate_service,
+                    overall_service = true,
+                    status = ?status,
+                    reason = "engine_unhealthy",
+                    "marking gRPC health services as not serving"
+                ),
+                Err(error) => warn!(
+                    %error,
+                    generate_service,
+                    overall_service = true,
+                    status = ?status,
+                    reason = "health_channel_closed",
+                    "engine health channel closed; marking gRPC health services as not serving"
+                ),
+            }
+            true
+        }
+        _ = shutdown.cancelled() => {
+            info!(
+                generate_service,
+                overall_service = true,
+                status = ?status,
+                reason = "server_shutdown",
+                "server shutting down; marking gRPC health services as not serving"
+            );
+            false
+        }
+    };
+
+    health_reporter.set_not_serving::<GenerateGrpcService>().await;
+    // Generate is currently the only engine-backed gRPC service, so overall
+    // server health intentionally mirrors it.
+    health_reporter.set_service_status("", status).await;
+
+    if health_event_first {
+        shutdown.cancelled().await;
+        info!(
+            generate_service,
+            overall_service = true,
+            reason = "server_shutdown",
+            "server shutting down; closing gRPC health watches"
+        );
+    }
+
+    health_reporter.clear_service_status(generate_service).await;
+    health_reporter.clear_service_status("").await;
+}
@@ -4,6 +4,7 @@
 //! gRPC Generate service backed by the shared [`vllm_text::TextLlm`] facade.

 mod convert;
+mod health;

 use std::pin::Pin;
 use std::sync::Arc;
@@ -24,8 +25,11 @@ pub mod pb {
    tonic::include_proto!("vllm");
 }

+pub(crate) use health::monitor_health;
 pub use pb::generate_server::GenerateServer;

+pub(crate) type GenerateGrpcService = GenerateServer<GenerateServiceImpl>;
+
 #[cfg(test)]
 mod tests;

@@ -16,6 +16,10 @@ use tokio::io::{AsyncReadExt as _, AsyncWriteExt as _};
 use tokio::net::TcpStream;
 use tokio_openssl::SslStream;
 use tonic::transport::{Channel, Endpoint, Server as TonicServer, Uri};
+use tonic_health::pb::HealthCheckRequest;
+use tonic_health::pb::health_check_response::ServingStatus as HealthServingStatus;
+use tonic_health::pb::health_client::HealthClient;
+use tonic_health::server::health_reporter;
 use tower::service_fn;
 use vllm_chat::{
    ChatBackend, ChatLlm, ChatRenderer, ChatRequest, ChatTextBackend, DefaultChatOutputProcessor,
@@ -200,7 +204,11 @@ impl ChatRenderer for FakeTextBackend {
 async fn setup_grpc_service(
    engine_id: impl Into<EngineId>,
    output_specs: Vec<(Vec<u32>, Option<EngineCoreFinishReason>)>,
-) -> (GenerateServer<GenerateServiceImpl>, MockEngineTask) {
+) -> (
+    GenerateServer<GenerateServiceImpl>,
+    tokio::sync::watch::Receiver<bool>,
+    MockEngineTask,
+) {
    let ipc = IpcNamespace::new().expect("create ipc namespace");
    let handshake_address = ipc.handshake_endpoint();
    let engine_id = engine_id.into();
@@ -232,6 +240,7 @@ async fn setup_grpc_service(
    )
    .await
    .expect("connect client");
+    let engine_health = client.subscribe_health();

    let chat = ChatLlm::from_shared_backend(
        test_llm(client),
@@ -240,6 +249,7 @@ async fn setup_grpc_service(
    let state = Arc::new(AppState::new(vec!["test-model".to_string()], chat));
    (
        GenerateServer::new(GenerateServiceImpl::new(state)),
+        engine_health,
        engine_task,
    )
 }
@@ -254,25 +264,51 @@ async fn grpc_test_server(
    tokio::task::JoinHandle<()>,
    MockEngineTask,
 ) {
-    let (svc, engine_task) = setup_grpc_service(engine_id, output_specs).await;
+    let (svc, engine_health, engine_task) = setup_grpc_service(engine_id, output_specs).await;
+    let (channel, server_task) = start_grpc_test_server(
+        svc,
+        engine_health,
+        tokio_util::sync::CancellationToken::new(),
+    )
+    .await;
+    (GenerateClient::new(channel), server_task, engine_task)
+}
+
+async fn start_grpc_test_server(
+    generate_service: GenerateServer<GenerateServiceImpl>,
+    engine_health: tokio::sync::watch::Receiver<bool>,
+    shutdown: tokio_util::sync::CancellationToken,
+) -> (Channel, tokio::task::JoinHandle<()>) {
+    let (health_reporter, health_service) = health_reporter();
+    health_reporter.set_serving::<GenerateServer<GenerateServiceImpl>>().await;

    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.expect("bind grpc listener");
    let addr = listener.local_addr().expect("local addr");

    let server_task = tokio::spawn(async move {
        let incoming = MaybeTlsListener::plain(Listener::Tcp(listener));
-        TonicServer::builder()
-            .add_service(svc)
-            .serve_with_incoming(incoming)
-            .await
-            .expect("grpc server");
+        let server = TonicServer::builder()
+            .add_service(health_service)
+            .add_service(generate_service)
+            .serve_with_incoming_shutdown(incoming, shutdown.clone().cancelled_owned());
+        let health_monitor =
+            super::monitor_health(health_reporter, engine_health, shutdown.clone());
+        let server = async move {
+            let result = server.await;
+            shutdown.cancel();
+            result
+        };
+        let (server_result, ()) = tokio::join!(server, health_monitor);
+        server_result.expect("grpc server");
    });

-    let grpc_client = GenerateClient::connect(format!("http://{addr}"))
+    let channel = Endpoint::from_shared(format!("http://{addr}"))
+        .expect("grpc endpoint")
+        .connect()
        .await
-        .expect("connect grpc client");
+        .expect("connect grpc channel");

-    (grpc_client, server_task, engine_task)
+    (channel, server_task)
 }

 /// Spin up a TLS gRPC server (server cert from `certs`, `cert_reqs` mTLS mode).
@@ -283,7 +319,7 @@ async fn grpc_tls_test_server(
    certs: &TestCerts,
    cert_reqs: i32,
 ) -> (String, tokio::task::JoinHandle<()>, MockEngineTask) {
-    let (svc, engine_task) = setup_grpc_service(engine_id, output_specs).await;
+    let (svc, _engine_health, engine_task) = setup_grpc_service(engine_id, output_specs).await;
    let context = tls::build_grpc_server_config(&server_tls(certs, cert_reqs))
        .expect("build grpc tls config");

@@ -373,7 +409,8 @@ async fn grpc_server_with_keepalive(
    engine_id: impl Into<EngineId>,
    keepalive: Option<Duration>,
 ) -> (String, tokio::task::JoinHandle<()>, MockEngineTask) {
-    let (svc, engine_task) = setup_grpc_service(engine_id, default_stream_output_specs()).await;
+    let (svc, _engine_health, engine_task) =
+        setup_grpc_service(engine_id, default_stream_output_specs()).await;

    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.expect("bind grpc listener");
    let addr = listener.local_addr().expect("local addr").to_string();
@@ -1035,3 +1072,129 @@ async fn grpc_without_keepalive_keeps_unresponsive_connection_open() {

    server_task.abort();
 }
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+#[serial]
+async fn grpc_health_transitions_to_not_serving_when_engine_becomes_unhealthy() {
+    let (generate_service, _connected_engine_health, _engine_task) =
+        setup_grpc_service(b"engine-grpc-health-failure", default_stream_output_specs()).await;
+    let (engine_health_tx, engine_health) = tokio::sync::watch::channel(true);
+    let (channel, server_task) = start_grpc_test_server(
+        generate_service,
+        engine_health,
+        tokio_util::sync::CancellationToken::new(),
+    )
+    .await;
+    let mut health_client = HealthClient::new(channel);
+
+    let mut health_streams = Vec::new();
+    for service in ["vllm.Generate", ""] {
+        let service_label = if service.is_empty() {
+            "overall"
+        } else {
+            service
+        };
+        let mut stream = health_client
+            .watch(HealthCheckRequest {
+                service: service.to_string(),
+            })
+            .await
+            .unwrap_or_else(|error| {
+                panic!("failed to start health watch for {service_label}: {error}")
+            })
+            .into_inner();
+        let initial = stream
+            .message()
+            .await
+            .unwrap_or_else(|error| {
+                panic!("failed to read initial health status for {service_label}: {error}")
+            })
+            .unwrap_or_else(|| {
+                panic!("health watch for {service_label} ended before its initial status")
+            });
+        assert_eq!(
+            initial.status,
+            HealthServingStatus::Serving as i32,
+            "unexpected initial health status for {service_label}"
+        );
+        health_streams.push((service_label, stream));
+    }
+
+    engine_health_tx.send(false).expect("publish unhealthy engine state");
+
+    for (service_label, mut stream) in health_streams {
+        let update = tokio::time::timeout(Duration::from_secs(2), stream.message())
+            .await
+            .unwrap_or_else(|_| panic!("timed out waiting for health update for {service_label}"))
+            .unwrap_or_else(|error| {
+                panic!("failed to read health update for {service_label}: {error}")
+            })
+            .unwrap_or_else(|| panic!("health watch for {service_label} ended before its update"));
+        assert_eq!(
+            update.status,
+            HealthServingStatus::NotServing as i32,
+            "unexpected health status for {service_label}"
+        );
+    }
+
+    server_task.abort();
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+#[serial]
+async fn grpc_health_watch_closes_on_graceful_shutdown() {
+    let (generate_service, engine_health, _engine_task) = setup_grpc_service(
+        b"engine-grpc-health-shutdown",
+        default_stream_output_specs(),
+    )
+    .await;
+    let shutdown = tokio_util::sync::CancellationToken::new();
+    let (channel, server_task) =
+        start_grpc_test_server(generate_service, engine_health, shutdown.clone()).await;
+    let mut health_client = HealthClient::new(channel);
+    let mut stream = health_client
+        .watch(HealthCheckRequest {
+            service: "vllm.Generate".to_string(),
+        })
+        .await
+        .expect("start health watch for vllm.Generate")
+        .into_inner();
+
+    let initial = stream
+        .message()
+        .await
+        .expect("read initial health status for vllm.Generate")
+        .expect("health watch ended before its initial status");
+    assert_eq!(
+        initial.status,
+        HealthServingStatus::Serving as i32,
+        "unexpected initial health status for vllm.Generate"
+    );
+
+    shutdown.cancel();
+
+    let update = tokio::time::timeout(Duration::from_secs(2), stream.message())
+        .await
+        .expect("timed out waiting for shutdown health update for vllm.Generate")
+        .expect("failed to read shutdown health update for vllm.Generate")
+        .expect("health watch ended before its shutdown update");
+    assert_eq!(
+        update.status,
+        HealthServingStatus::NotServing as i32,
+        "unexpected shutdown health status for vllm.Generate"
+    );
+
+    let stream_end = tokio::time::timeout(Duration::from_secs(2), stream.message())
+        .await
+        .expect("timed out waiting for vllm.Generate health watch to close")
+        .expect("failed while closing vllm.Generate health watch");
+    assert!(
+        stream_end.is_none(),
+        "vllm.Generate health watch remained open"
+    );
+
+    tokio::time::timeout(Duration::from_secs(2), server_task)
+        .await
+        .expect("timed out waiting for gRPC server shutdown")
+        .expect("gRPC server task failed");
+}
@@ -39,6 +39,7 @@ use tokio::net::TcpListener;
 use tokio::time::{Instant, sleep_until};
 use tokio_util::sync::CancellationToken;
 use tonic::transport::Server as TonicServer;
+use tonic_health::server::health_reporter;
 use tower::ServiceExt as _;
 use tracing::{info, trace, warn};
 use vllm_chat::{ChatLlm, LoadModelBackendsOptions, load_model_backends};
@@ -203,14 +204,19 @@ where
            .map(tls::build_grpc_server_config)
            .transpose()
            .context("invalid gRPC TLS configuration")?;
-        let svc = grpc::GenerateServer::new(grpc::GenerateServiceImpl::new(state.clone()));
+        let (health_reporter, health_service) = health_reporter();
+        let engine_health = state.engine_core_client().subscribe_health();
+        health_reporter.set_serving::<grpc::GenerateGrpcService>().await;
+        let generate_service =
+            grpc::GenerateGrpcService::new(grpc::GenerateServiceImpl::new(state.clone()));
        let svc = TonicServer::builder()
            .http2_keepalive_interval(Some(GRPC_KEEPALIVE_INTERVAL))
            .http2_keepalive_timeout(Some(GRPC_KEEPALIVE_TIMEOUT))
            .layer(middleware::request_runtime_layer(state.clone()))
-            .add_service(svc);
+            .add_service(health_service)
+            .add_service(generate_service);
        info!(%addr, tls = grpc_tls.is_some(), "starting gRPC server");
-        Some((grpc_listener, svc, grpc_tls))
+        Some((grpc_listener, svc, grpc_tls, health_reporter, engine_health))
    } else {
        None
    };
@@ -294,7 +300,8 @@ where
        let server_shutdown = server_shutdown.clone();
        let force_shutdown = force_shutdown.clone();
        async move {
-            let Some((grpc_listener, svc, grpc_tls)) = grpc_setup else {
+            let Some((grpc_listener, svc, grpc_tls, health_reporter, engine_health)) = grpc_setup
+            else {
                // No gRPC configured: just wait for shutdown so we do not race the
                // join! by resolving early and tripping the cancellation token.
                shutdown.cancelled().await;
@@ -304,19 +311,26 @@ where
                Some(context) => MaybeTlsListener::tls(grpc_listener, context),
                None => MaybeTlsListener::plain(grpc_listener),
            };
-            let server = svc.serve_with_incoming_shutdown(incoming, shutdown.cancelled_owned());
+            let server =
+                svc.serve_with_incoming_shutdown(incoming, shutdown.clone().cancelled_owned());
+            let health_monitor = grpc::monitor_health(health_reporter, engine_health, shutdown);

-            let result = tokio::select! {
-                result = server => {
-                    result.context("gRPC server failed")
-                }
-                _ = force_shutdown.cancelled() => {
-                    warn!("gRPC graceful shutdown deadline elapsed; aborting server");
-                    Ok(())
-                }
+            let server = async move {
+                let result = tokio::select! {
+                    result = server => {
+                        result.context("gRPC server failed")
+                    }
+                    _ = force_shutdown.cancelled() => {
+                        warn!("gRPC graceful shutdown deadline elapsed; aborting server");
+                        Ok(())
+                    }
+                };
+
+                server_shutdown.cancel();
+                result
            };

-            server_shutdown.cancel();
+            let (result, ()) = tokio::join!(server, health_monitor);
            result
        }
    };
@@ -60,6 +60,8 @@ def _make_quick_allreduce(
    qar.use_fp16_kernels = use_fp16_kernels
    qar.qr_quant_level = QuickReduceRegime[quant_level]
    qar.qr_max_size = qr_max_size
+    qar.qr_min_size = None
+    qar.qr_quantization_min_size = None
    return qar


@@ -511,13 +513,21 @@ def test_quick_reduce_regime_values():
    assert QuickReduceRegime.INT8.value == 1
    assert QuickReduceRegime.INT6.value == 2
    assert QuickReduceRegime.INT4.value == 3
-    assert QuickReduceRegime.NONE.value == 4
+    assert QuickReduceRegime.INT3.value == 4
+    assert QuickReduceRegime.NONE.value == 5


 def test_quick_reduce_regime_names():
    from vllm.distributed.device_communicators.quick_all_reduce import QuickReduceRegime

-    assert set(QuickReduceRegime.__members__) == {"FP", "INT8", "INT6", "INT4", "NONE"}
+    assert set(QuickReduceRegime.__members__) == {
+        "FP",
+        "INT8",
+        "INT6",
+        "INT4",
+        "INT3",
+        "NONE",
+    }


@pytest.mark.parametrize("quant_level", QUANT_LEVELS + ["NONE"])
@@ -693,7 +703,7 @@ def test_quick_allreduce_min_size_table():
    for dtype in [torch.float16, torch.bfloat16]:
        for world_size in QuickAllReduce._SUPPORTED_WORLD_SIZES:
            min_sizes = QuickAllReduce._QR_MIN_SIZE[(dtype, world_size)]
-            assert len(min_sizes) == 4
+            assert len(min_sizes) == 5
            assert all(size > 0 for size in min_sizes)


@@ -0,0 +1,880 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""End-to-end tests for routed-expert capture on the monolithic MoE path.
+
+These tests exercise the wiring that lets ``RoutedExpertsCapturer`` see the
+expert IDs picked by FlashInfer's fused router-and-experts kernels (the
+"monolithic" path). When ``set_capture_fn`` is installed on
+a ``FusedMoEExpertsMonolithic`` subclass that supports it, the kernel call
+should:
+
+  * allocate an int16 ``(num_tokens, top_k)`` buffer,
+  * pass it to FlashInfer as ``routing_replay_out``,
+  * invoke the callback after the kernel returns.
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from vllm.model_executor.layers.fused_moe.activation import MoEActivation
+from vllm.model_executor.layers.fused_moe.config import (
+    FusedMoEConfig,
+    FusedMoEParallelConfig,
+    FusedMoEQuantConfig,
+    RoutingMethodType,
+    fp8_w8a8_moe_quant_config,
+)
+from vllm.model_executor.layers.fused_moe.experts.trtllm_bf16_moe import (
+    TrtLlmBf16ExpertsMonolithic,
+)
+from vllm.model_executor.layers.fused_moe.experts.trtllm_fp8_moe import (
+    TrtLlmFp8ExpertsMonolithic,
+)
+from vllm.model_executor.layers.fused_moe.experts.trtllm_nvfp4_moe import (
+    TrtLlmNvFp4ExpertsMonolithic,
+)
+from vllm.platforms import current_platform
+
+try:
+    from vllm.utils.flashinfer import has_flashinfer_trtllm_fused_moe
+except ImportError:
+    pytest.skip("flashinfer not available", allow_module_level=True)
+
+if not has_flashinfer_trtllm_fused_moe() or not current_platform.is_cuda():
+    pytest.skip(
+        "Requires FlashInfer TRT-LLM fused MoE on CUDA",
+        allow_module_level=True,
+    )
+
+if not current_platform.has_device_capability(100):
+    pytest.skip(
+        "TRT-LLM fused MoE kernels require SM100+",
+        allow_module_level=True,
+    )
+
+
+def _shuffle_bf16_weights_block_major_k(
+    w: torch.Tensor, epilogue_tile_m: int = 64, block_k: int = 128
+) -> torch.Tensor:
+    """Reshape ``w`` (E, M, K) into the ``BlockMajorK`` layout expected by
+    ``trtllm_bf16_moe``: ``(E, K/block_k, M, block_k)`` after a per-expert
+    row shuffle.
+    """
+    from flashinfer import shuffle_matrix_a
+    from flashinfer.fused_moe import convert_to_block_layout
+
+    num_experts = w.shape[0]
+    shuffled = []
+    for i in range(num_experts):
+        t = shuffle_matrix_a(w[i].view(torch.uint8), epilogue_tile_m)
+        shuffled.append(convert_to_block_layout(t, block_k))
+    return torch.stack(shuffled).view(torch.bfloat16)
+
+
+def _make_bf16_monolithic_experts(
+    num_experts: int,
+    top_k: int,
+    hidden_size: int,
+    intermediate_size: int,
+    routing_method: RoutingMethodType,
+    device: torch.device,
+) -> tuple[TrtLlmBf16ExpertsMonolithic, torch.Tensor, torch.Tensor]:
+    """Construct the monolithic BF16 experts plus the BlockMajorK weights
+    expected by ``trtllm_bf16_moe``.
+    """
+    parallel_cfg = FusedMoEParallelConfig.make_no_parallel()
+    moe_config = FusedMoEConfig(
+        num_experts=num_experts,
+        experts_per_token=top_k,
+        hidden_dim=hidden_size,
+        intermediate_size=intermediate_size,
+        num_local_experts=num_experts,
+        num_logical_experts=num_experts,
+        moe_parallel_config=parallel_cfg,
+        in_dtype=torch.bfloat16,
+        activation=MoEActivation.SILU,
+        device=device,
+        routing_method=routing_method,
+        max_num_tokens=max(8, 1),
+    )
+    quant_config = FusedMoEQuantConfig.make(
+        quant_dtype=None,
+        per_act_token_quant=False,
+        per_out_ch_quant=False,
+        block_shape=None,
+    )
+
+    experts = TrtLlmBf16ExpertsMonolithic(
+        moe_config=moe_config, quant_config=quant_config
+    )
+
+    gemm1 = (
+        torch.randn(
+            num_experts,
+            2 * intermediate_size,
+            hidden_size,
+            device=device,
+            dtype=torch.bfloat16,
+        )
+        * 0.1
+    )
+    gemm2 = (
+        torch.randn(
+            num_experts,
+            hidden_size,
+            intermediate_size,
+            device=device,
+            dtype=torch.bfloat16,
+        )
+        * 0.1
+    )
+    w13 = _shuffle_bf16_weights_block_major_k(gemm1)
+    w2 = _shuffle_bf16_weights_block_major_k(gemm2)
+    return experts, w13, w2
+
+
+def _run_bf16_monolithic(
+    experts: TrtLlmBf16ExpertsMonolithic,
+    hidden_states: torch.Tensor,
+    w13: torch.Tensor,
+    w2: torch.Tensor,
+    router_logits: torch.Tensor,
+    num_experts: int,
+    *,
+    n_group: int | None = None,
+    topk_group: int | None = None,
+    routed_scaling_factor: float | None = None,
+    e_score_correction_bias: torch.Tensor | None = None,
+) -> torch.Tensor:
+    return experts.apply(
+        hidden_states=hidden_states,
+        w1=w13,
+        w2=w2,
+        router_logits=router_logits,
+        activation=MoEActivation.SILU,
+        global_num_experts=num_experts,
+        expert_map=None,
+        a1q_scale=None,
+        apply_router_weight_on_input=False,
+        num_expert_group=n_group,
+        topk_group=topk_group,
+        e_score_correction_bias=e_score_correction_bias,
+        routed_scaling_factor=routed_scaling_factor,
+    )
+
+
+_DSV3_NUM_EXPERTS = 32
+_DSV3_N_GROUP = 4
+_DSV3_TOPK_GROUP = 2
+
+
+def _make_dsv3_routing_bias(num_experts: int, device: torch.device) -> torch.Tensor:
+    return torch.randn(num_experts, device=device, dtype=torch.bfloat16)
+
+
+@pytest.mark.parametrize("num_tokens", [2, 7, 16])
+@pytest.mark.parametrize("top_k", [2, 4])
+def test_trtllm_bf16_monolithic_routing_replay_records_valid_experts(
+    num_tokens: int,
+    top_k: int,
+) -> None:
+    """The capture callback should receive the int16 routed-expert IDs the
+    kernel actually used, the values should be valid expert indices, and
+    each token should pick ``top_k`` distinct experts."""
+    if top_k > _DSV3_N_GROUP * _DSV3_TOPK_GROUP:
+        pytest.skip(
+            f"DSV3 requires top_k <= n_group * topk_group "
+            f"({_DSV3_N_GROUP * _DSV3_TOPK_GROUP})"
+        )
+    torch.manual_seed(0)
+    device = torch.device("cuda:0")
+
+    num_experts = _DSV3_NUM_EXPERTS
+    hidden_size = 1024
+    intermediate_size = 1024
+
+    experts, w13, w2 = _make_bf16_monolithic_experts(
+        num_experts=num_experts,
+        top_k=top_k,
+        hidden_size=hidden_size,
+        intermediate_size=intermediate_size,
+        routing_method=RoutingMethodType.DeepSeekV3,
+        device=device,
+    )
+
+    captured: list[torch.Tensor] = []
+
+    def capture_fn(replay_out: torch.Tensor) -> None:
+        captured.append(replay_out.clone())
+
+    assert experts.supports_routing_replay_capture()
+    experts.set_capture_fn(capture_fn)
+
+    hidden_states = (
+        torch.randn(num_tokens, hidden_size, device=device, dtype=torch.bfloat16) * 0.1
+    )
+    router_logits = torch.rand(
+        num_tokens, num_experts, device=device, dtype=torch.float32
+    )
+    routing_bias = _make_dsv3_routing_bias(num_experts, device)
+
+    _ = _run_bf16_monolithic(
+        experts,
+        hidden_states=hidden_states,
+        w13=w13,
+        w2=w2,
+        router_logits=router_logits,
+        num_experts=num_experts,
+        n_group=_DSV3_N_GROUP,
+        topk_group=_DSV3_TOPK_GROUP,
+        routed_scaling_factor=1.0,
+        e_score_correction_bias=routing_bias,
+    )
+
+    assert len(captured) == 1
+    replay = captured[0]
+    assert replay.dtype == torch.int16
+    assert replay.shape == (num_tokens, top_k)
+    assert (replay >= 0).all(), f"got out-of-range values: {replay}"
+    assert (replay < num_experts).all(), f"got out-of-range values: {replay}"
+
+    for t in range(num_tokens):
+        unique = replay[t].unique()
+        assert unique.numel() == top_k, (
+            f"token {t}: expected {top_k} distinct experts, "
+            f"got {unique.numel()} ({replay[t].tolist()})"
+        )
+
+
+@pytest.mark.parametrize("num_tokens", [2, 7, 16])
+@pytest.mark.parametrize(
+    "routing_method",
+    [
+        RoutingMethodType.Renormalize,
+        RoutingMethodType.RenormalizeNaive,
+    ],
+)
+def test_trtllm_bf16_monolithic_routing_replay_non_dsv3(
+    num_tokens: int,
+    routing_method: RoutingMethodType,
+) -> None:
+    """Routing replay works for non-DeepSeekV3 routing methods too.
+    FlashInfer's ``routing_replay_out`` is routing-method-agnostic."""
+    torch.manual_seed(0)
+    device = torch.device("cuda:0")
+
+    num_experts = 8
+    top_k = 2
+    hidden_size = 1024
+    intermediate_size = 1024
+
+    experts, w13, w2 = _make_bf16_monolithic_experts(
+        num_experts=num_experts,
+        top_k=top_k,
+        hidden_size=hidden_size,
+        intermediate_size=intermediate_size,
+        routing_method=routing_method,
+        device=device,
+    )
+
+    captured: list[torch.Tensor] = []
+    experts.set_capture_fn(lambda r: captured.append(r.clone()))
+
+    hidden_states = (
+        torch.randn(num_tokens, hidden_size, device=device, dtype=torch.bfloat16) * 0.1
+    )
+    router_logits = torch.rand(
+        num_tokens, num_experts, device=device, dtype=torch.float32
+    )
+
+    _ = _run_bf16_monolithic(
+        experts,
+        hidden_states=hidden_states,
+        w13=w13,
+        w2=w2,
+        router_logits=router_logits,
+        num_experts=num_experts,
+    )
+
+    assert len(captured) == 1
+    replay = captured[0]
+    assert replay.dtype == torch.int16
+    assert replay.shape == (num_tokens, top_k)
+    assert (replay >= 0).all(), f"got out-of-range values: {replay}"
+    assert (replay < num_experts).all(), f"got out-of-range values: {replay}"
+    for t in range(num_tokens):
+        unique = replay[t].unique()
+        assert unique.numel() == top_k, (
+            f"token {t}: expected {top_k} distinct experts, "
+            f"got {unique.numel()} ({replay[t].tolist()})"
+        )
+
+
+def test_trtllm_bf16_monolithic_capture_disabled_skips_buffer_alloc() -> None:
+    """With no callback installed the kernel should not see a
+    ``routing_replay_out`` tensor — verify the helper short-circuits."""
+    torch.manual_seed(0)
+    device = torch.device("cuda:0")
+    experts, _, _ = _make_bf16_monolithic_experts(
+        num_experts=_DSV3_NUM_EXPERTS,
+        top_k=2,
+        hidden_size=1024,
+        intermediate_size=1024,
+        routing_method=RoutingMethodType.DeepSeekV3,
+        device=device,
+    )
+    # No callback installed.
+    buf = experts._maybe_make_routing_replay_buffer(num_tokens=4, device=device)
+    assert buf is None
+
+    # Dispatch is also a no-op.
+    experts._maybe_dispatch_routing_replay(buf, num_tokens=4)
+
+
+def test_trtllm_bf16_monolithic_supports_capture_for_all_routing() -> None:
+    """FlashInfer's ``routing_replay_out`` is supported by all routing
+    methods, so ``supports_routing_replay_capture`` should be True
+    regardless of routing method."""
+    device = torch.device("cuda:0")
+    for routing_method in (
+        RoutingMethodType.DeepSeekV3,
+        RoutingMethodType.Renormalize,
+        RoutingMethodType.RenormalizeNaive,
+    ):
+        experts, _, _ = _make_bf16_monolithic_experts(
+            num_experts=_DSV3_NUM_EXPERTS,
+            top_k=2,
+            hidden_size=1024,
+            intermediate_size=1024,
+            routing_method=routing_method,
+            device=device,
+        )
+        assert experts.supports_routing_replay_capture() is True, (
+            f"{routing_method!r} should support routing replay capture"
+        )
+
+
+def test_trtllm_bf16_monolithic_capture_buffer_shape_and_dtype() -> None:
+    """When capture is installed, the allocated buffer is int16 and shaped
+    ``(num_tokens, experts_per_token)``."""
+    device = torch.device("cuda:0")
+    experts, _, _ = _make_bf16_monolithic_experts(
+        num_experts=_DSV3_NUM_EXPERTS,
+        top_k=4,
+        hidden_size=1024,
+        intermediate_size=1024,
+        routing_method=RoutingMethodType.DeepSeekV3,
+        device=device,
+    )
+    experts.set_capture_fn(lambda r: None)
+    buf = experts._maybe_make_routing_replay_buffer(num_tokens=11, device=device)
+    assert buf is not None
+    assert buf.dtype == torch.int16
+    assert buf.shape[0] >= 11
+    assert buf.shape[1] == 4
+    assert buf.device.type == "cuda"
+
+
+def test_routed_experts_capturer_e2e_via_monolithic_experts() -> None:
+    """End-to-end: bind ``RoutedExpertsCapturer.capture`` as the callback
+    on the monolithic experts and verify the captured rows land in the
+    capturer's device buffer at the correct layer slot.
+
+    Mirrors the wiring done in ``GPUModelRunner._bind_routed_experts_capturer``
+    for the monolithic path: a single closure is installed on the monolithic
+    ``fused_experts`` (in addition to ``router.set_capture_fn`` on the
+    non-monolithic path) and the capturer routes per-layer based on the
+    closed-over ``layer_id``.
+    """
+    from vllm.model_executor.layers.fused_moe.routed_experts_capturer import (
+        RoutedExpertsCapturer,
+    )
+
+    torch.manual_seed(7)
+    device = torch.device("cuda:0")
+    num_tokens = 4
+    top_k = 2
+    num_experts = _DSV3_NUM_EXPERTS
+    hidden_size = 1024
+    intermediate_size = 1024
+
+    experts, w13, w2 = _make_bf16_monolithic_experts(
+        num_experts=num_experts,
+        top_k=top_k,
+        hidden_size=hidden_size,
+        intermediate_size=intermediate_size,
+        routing_method=RoutingMethodType.DeepSeekV3,
+        device=device,
+    )
+
+    num_layers = 3
+    layer_id = 1
+    capturer = RoutedExpertsCapturer.__new__(RoutedExpertsCapturer)
+    capturer.dp_rank = 0
+    capturer.tp_size = 1
+    capturer.device_buffer = torch.full(
+        (num_tokens + 4, num_layers, top_k),
+        -1,
+        dtype=torch.int32,
+        device=device,
+    )
+
+    def capture_fn(replay_out: torch.Tensor) -> None:
+        capturer.capture(layer_id, replay_out)
+
+    experts.set_capture_fn(capture_fn)
+
+    hidden_states = (
+        torch.randn(num_tokens, hidden_size, device=device, dtype=torch.bfloat16) * 0.1
+    )
+    router_logits = torch.rand(
+        num_tokens, num_experts, device=device, dtype=torch.float32
+    )
+    routing_bias = _make_dsv3_routing_bias(num_experts, device)
+
+    # Patch get_forward_context to return a dp_metadata=None context so the
+    # capturer takes the single-DP branch.
+    import vllm.model_executor.layers.fused_moe.routed_experts_capturer as rec
+
+    with patch.object(
+        rec,
+        "get_forward_context",
+        return_value=SimpleNamespace(dp_metadata=None),
+    ):
+        _ = _run_bf16_monolithic(
+            experts,
+            hidden_states=hidden_states,
+            w13=w13,
+            w2=w2,
+            router_logits=router_logits,
+            num_experts=num_experts,
+            n_group=_DSV3_N_GROUP,
+            topk_group=_DSV3_TOPK_GROUP,
+            routed_scaling_factor=1.0,
+            e_score_correction_bias=routing_bias,
+        )
+
+    captured = capturer.device_buffer[:num_tokens, layer_id, :].cpu()
+    # Valid expert IDs at this layer.
+    assert (captured >= 0).all()
+    assert (captured < num_experts).all()
+    for t in range(num_tokens):
+        unique = captured[t].unique()
+        assert unique.numel() == top_k, (
+            f"token {t}: expected {top_k} distinct experts at layer "
+            f"{layer_id}, got {unique.numel()}"
+        )
+
+    # Other layers / trailing token rows untouched.
+    for other_layer in range(num_layers):
+        if other_layer == layer_id:
+            continue
+        assert (capturer.device_buffer[:, other_layer, :].cpu() == -1).all(), (
+            f"layer {other_layer} should be untouched, got writes"
+        )
+    assert (capturer.device_buffer[num_tokens:, layer_id, :].cpu() == -1).all(), (
+        "tail rows beyond num_tokens should remain sentinel"
+    )
+
+
+# ----------------------------------------------------------------------------
+# FP8 block-scale (DeepSeekFp8) — vLLM's ``TrtLlmFp8ExpertsMonolithic``
+# ----------------------------------------------------------------------------
+
+
+def _make_fp8_block_scale_monolithic_experts(
+    num_experts: int,
+    top_k: int,
+    hidden_size: int,
+    intermediate_size: int,
+    device: torch.device,
+) -> tuple[TrtLlmFp8ExpertsMonolithic, torch.Tensor, torch.Tensor]:
+    """Set up ``TrtLlmFp8ExpertsMonolithic`` for the DeepSeekFp8 block-scale
+    code path with DSV3 routing.
+
+    Weights are shuffled into the BlockMajorK layout the kernel expects
+    (same helper the vLLM weight loader uses for DeepSeek-FP8 models).
+    """
+    from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
+        _shuffle_deepseek_fp8_moe_weights,
+    )
+
+    block_k = 128
+    parallel_cfg = FusedMoEParallelConfig.make_no_parallel()
+    moe_config = FusedMoEConfig(
+        num_experts=num_experts,
+        experts_per_token=top_k,
+        hidden_dim=hidden_size,
+        intermediate_size=intermediate_size,
+        num_local_experts=num_experts,
+        num_logical_experts=num_experts,
+        moe_parallel_config=parallel_cfg,
+        in_dtype=torch.bfloat16,
+        activation=MoEActivation.SILU,
+        device=device,
+        routing_method=RoutingMethodType.DeepSeekV3,
+        max_num_tokens=max(8, 1),
+    )
+
+    # Random fp8 weights + ones-block scales (the kernel decoder only cares
+    # that the per-block scales are present and finite for routing/replay).
+    gemm1 = torch.randn(
+        num_experts, 2 * intermediate_size, hidden_size, device=device
+    ).to(torch.float8_e4m3fn)
+    gemm2 = torch.randn(num_experts, hidden_size, intermediate_size, device=device).to(
+        torch.float8_e4m3fn
+    )
+    w13_shuffled, w2_shuffled = _shuffle_deepseek_fp8_moe_weights(gemm1, gemm2)
+
+    w1_scale = torch.ones(
+        num_experts,
+        2 * intermediate_size // block_k,
+        hidden_size // block_k,
+        device=device,
+        dtype=torch.float32,
+    )
+    w2_scale = torch.ones(
+        num_experts,
+        hidden_size // block_k,
+        intermediate_size // block_k,
+        device=device,
+        dtype=torch.float32,
+    )
+    quant_config = fp8_w8a8_moe_quant_config(
+        w1_scale=w1_scale,
+        w2_scale=w2_scale,
+        block_shape=[block_k, block_k],
+        per_act_token_quant=False,
+    )
+
+    experts = TrtLlmFp8ExpertsMonolithic(
+        moe_config=moe_config, quant_config=quant_config
+    )
+    return experts, w13_shuffled, w2_shuffled
+
+
+def _run_fp8_block_scale_monolithic(
+    experts: TrtLlmFp8ExpertsMonolithic,
+    hidden_states_fp8: torch.Tensor,
+    hidden_states_scale: torch.Tensor,
+    w13: torch.Tensor,
+    w2: torch.Tensor,
+    router_logits: torch.Tensor,
+    num_experts: int,
+    routing_bias: torch.Tensor,
+) -> torch.Tensor:
+    return experts.apply(
+        hidden_states=hidden_states_fp8,
+        w1=w13,
+        w2=w2,
+        router_logits=router_logits,
+        activation=MoEActivation.SILU,
+        global_num_experts=num_experts,
+        expert_map=None,
+        # The block-scale apply path reads ``a1q_scale`` and transposes it
+        # to ``(hidden_size/128, num_tokens)`` for the kernel call.
+        a1q_scale=hidden_states_scale,
+        apply_router_weight_on_input=False,
+        num_expert_group=_DSV3_N_GROUP,
+        topk_group=_DSV3_TOPK_GROUP,
+        e_score_correction_bias=routing_bias,
+        routed_scaling_factor=1.0,
+    )
+
+
+@pytest.mark.parametrize("num_tokens", [2, 7, 16])
+@pytest.mark.parametrize("top_k", [2, 4])
+def test_trtllm_fp8_block_scale_monolithic_routing_replay_records_valid_experts(
+    num_tokens: int,
+    top_k: int,
+) -> None:
+    """End-to-end: ``TrtLlmFp8ExpertsMonolithic`` (DeepSeekFp8 block-scale
+    path, DSV3 routing) captures valid expert IDs."""
+    if top_k > _DSV3_N_GROUP * _DSV3_TOPK_GROUP:
+        pytest.skip(
+            f"DSV3 requires top_k <= n_group * topk_group "
+            f"({_DSV3_N_GROUP * _DSV3_TOPK_GROUP})"
+        )
+    torch.manual_seed(0)
+    device = torch.device("cuda:0")
+
+    num_experts = _DSV3_NUM_EXPERTS
+    hidden_size = 1024
+    intermediate_size = 1024
+
+    experts, w13, w2 = _make_fp8_block_scale_monolithic_experts(
+        num_experts=num_experts,
+        top_k=top_k,
+        hidden_size=hidden_size,
+        intermediate_size=intermediate_size,
+        device=device,
+    )
+    assert experts.supports_routing_replay_capture()
+
+    captured: list[torch.Tensor] = []
+    experts.set_capture_fn(lambda r: captured.append(r.clone()))
+
+    # Per-token / per-block hidden scales (ones is fine for the routing
+    # path; the GEMM output isn't being asserted on).
+    hidden_states = (
+        torch.randn(num_tokens, hidden_size, device=device, dtype=torch.bfloat16) * 0.1
+    ).to(torch.float8_e4m3fn)
+    hidden_states_scale = torch.ones(
+        num_tokens, hidden_size // 128, device=device, dtype=torch.float32
+    )
+    router_logits = torch.rand(
+        num_tokens, num_experts, device=device, dtype=torch.float32
+    )
+    routing_bias = _make_dsv3_routing_bias(num_experts, device)
+
+    _ = _run_fp8_block_scale_monolithic(
+        experts,
+        hidden_states_fp8=hidden_states,
+        hidden_states_scale=hidden_states_scale,
+        w13=w13,
+        w2=w2,
+        router_logits=router_logits,
+        num_experts=num_experts,
+        routing_bias=routing_bias,
+    )
+
+    assert len(captured) == 1
+    replay = captured[0]
+    assert replay.dtype == torch.int16
+    assert replay.shape == (num_tokens, top_k)
+    assert (replay >= 0).all(), f"got out-of-range values: {replay}"
+    assert (replay < num_experts).all(), f"got out-of-range values: {replay}"
+    for t in range(num_tokens):
+        unique = replay[t].unique()
+        assert unique.numel() == top_k, (
+            f"token {t}: expected {top_k} distinct experts, "
+            f"got {unique.numel()} ({replay[t].tolist()})"
+        )
+
+
+# ----------------------------------------------------------------------------
+# NVFP4 — vLLM's ``TrtLlmNvFp4ExpertsMonolithic``
+# ----------------------------------------------------------------------------
+
+
+def _make_nvfp4_monolithic_experts(
+    num_experts: int,
+    top_k: int,
+    hidden_size: int,
+    intermediate_size: int,
+    device: torch.device,
+) -> tuple[
+    TrtLlmNvFp4ExpertsMonolithic,
+    torch.Tensor,  # w13 (packed nvfp4 uint8)
+    torch.Tensor,  # w13 block-scale (fp8)
+    torch.Tensor,  # w2  (packed nvfp4 uint8)
+    torch.Tensor,  # w2  block-scale (fp8)
+    torch.Tensor,  # input global scale (per-tensor float32)
+]:
+    """Set up ``TrtLlmNvFp4ExpertsMonolithic`` with NVFP4-quantized weights
+    and DSV3 routing.
+
+    NVFP4 = per-block-of-16 fp4 with an fp8 scale, plus a per-tensor
+    "global" scale. We follow the layout in ``test_ocp_mx_moe.py`` /
+    ``flashinfer/tests/moe/test_trtllm_gen_routed_fused_moe.py``:
+      * weights: uint8 (packed fp4) ``(E, M, K//2)``
+      * weight scales: fp8 ``(E, M, K//16)``
+      * hidden states: uint8 (packed fp4) ``(N, K//2)``
+      * hidden state scales: fp8 ``(N, K//16)``
+    """
+    from flashinfer import fp4_quantize
+
+    block_size = 16
+    parallel_cfg = FusedMoEParallelConfig.make_no_parallel()
+    moe_config = FusedMoEConfig(
+        num_experts=num_experts,
+        experts_per_token=top_k,
+        hidden_dim=hidden_size,
+        intermediate_size=intermediate_size,
+        num_local_experts=num_experts,
+        num_logical_experts=num_experts,
+        moe_parallel_config=parallel_cfg,
+        in_dtype=torch.bfloat16,
+        activation=MoEActivation.SILU,
+        device=device,
+        routing_method=RoutingMethodType.DeepSeekV3,
+        max_num_tokens=max(8, 1),
+    )
+
+    gemm1 = torch.randn(
+        num_experts,
+        2 * intermediate_size,
+        hidden_size,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    gemm2 = torch.randn(
+        num_experts,
+        hidden_size,
+        intermediate_size,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    # Per-tensor weight scaling factor (used to build ``g1_alphas`` /
+    # ``g2_alphas`` below).
+    w_global_scale = torch.tensor(1.0, device=device)
+    # Per-tensor input scaling factor.
+    a_global_scale = torch.tensor(1.0, device=device)
+
+    w13_q, w13_scale = fp4_quantize(
+        gemm1,
+        w_global_scale,
+        block_size,
+        sf_use_ue8m0=False,
+        is_sf_swizzled_layout=False,
+    )
+    w13_scale = w13_scale.view(torch.float8_e4m3fn).reshape(
+        num_experts, 2 * intermediate_size, hidden_size // block_size
+    )
+    w2_q, w2_scale = fp4_quantize(
+        gemm2,
+        w_global_scale,
+        block_size,
+        sf_use_ue8m0=False,
+        is_sf_swizzled_layout=False,
+    )
+    w2_scale = w2_scale.view(torch.float8_e4m3fn).reshape(
+        num_experts, hidden_size, intermediate_size // block_size
+    )
+
+    # NVFP4 dq scale chain: g1_alphas = w1_scale_2 * a1_scale_2,
+    # g2_alphas = w2_scale_2 * a2_scale_2. The kernel multiplies by these.
+    g_alphas = torch.full((num_experts,), 1.0, device=device, dtype=torch.float32)
+    a2_gscale = torch.full((num_experts,), 1.0, device=device, dtype=torch.float32)
+
+    quant_config = FusedMoEQuantConfig.make(
+        quant_dtype="nvfp4",
+        per_act_token_quant=False,
+        per_out_ch_quant=False,
+        block_shape=None,
+        w1_scale=w13_scale,
+        w2_scale=w2_scale,
+        g1_alphas=g_alphas,
+        g2_alphas=g_alphas,
+        a1_gscale=a_global_scale,
+        a2_gscale=a2_gscale,
+    )
+
+    experts = TrtLlmNvFp4ExpertsMonolithic(
+        moe_config=moe_config, quant_config=quant_config
+    )
+    return experts, w13_q, w13_scale, w2_q, w2_scale, a_global_scale
+
+
+def _run_nvfp4_monolithic(
+    experts: TrtLlmNvFp4ExpertsMonolithic,
+    hidden_states_q: torch.Tensor,
+    hidden_states_scale: torch.Tensor,
+    router_logits: torch.Tensor,
+    num_experts: int,
+    routing_bias: torch.Tensor,
+) -> torch.Tensor:
+    """The monolithic NVFP4 apply expects packed fp4 hidden states + the
+    matching fp8 per-block scale stored in the ``a1q_scale`` slot."""
+    # Stash the weight tensors on the experts in the locations the apply()
+    # implementation reads from (it pulls them from quant_config / scales
+    # already; w1/w2 come in as args).
+    return experts.apply(
+        hidden_states=hidden_states_q,
+        w1=experts._w13_packed,
+        w2=experts._w2_packed,
+        router_logits=router_logits,
+        activation=MoEActivation.SILU,
+        global_num_experts=num_experts,
+        expert_map=None,
+        a1q_scale=hidden_states_scale,
+        apply_router_weight_on_input=False,
+        num_expert_group=_DSV3_N_GROUP,
+        topk_group=_DSV3_TOPK_GROUP,
+        e_score_correction_bias=routing_bias,
+        routed_scaling_factor=1.0,
+    )
+
+
+@pytest.mark.parametrize("num_tokens", [2, 7, 16])
+@pytest.mark.parametrize("top_k", [2, 4])
+def test_trtllm_nvfp4_monolithic_routing_replay_records_valid_experts(
+    num_tokens: int,
+    top_k: int,
+) -> None:
+    """End-to-end: ``TrtLlmNvFp4ExpertsMonolithic`` captures valid expert IDs
+    on the DSV3 routing path."""
+    if top_k > _DSV3_N_GROUP * _DSV3_TOPK_GROUP:
+        pytest.skip(
+            f"DSV3 requires top_k <= n_group * topk_group "
+            f"({_DSV3_N_GROUP * _DSV3_TOPK_GROUP})"
+        )
+    from flashinfer import fp4_quantize
+
+    torch.manual_seed(0)
+    device = torch.device("cuda:0")
+
+    num_experts = _DSV3_NUM_EXPERTS
+    hidden_size = 1024
+    intermediate_size = 1024
+    block_size = 16
+
+    experts, w13_q, _w13_s, w2_q, _w2_s, a_gs = _make_nvfp4_monolithic_experts(
+        num_experts=num_experts,
+        top_k=top_k,
+        hidden_size=hidden_size,
+        intermediate_size=intermediate_size,
+        device=device,
+    )
+    # The apply() reads w1/w2 from its args, but we keep them on the experts
+    # for convenience of the helper.
+    experts._w13_packed = w13_q
+    experts._w2_packed = w2_q
+
+    assert experts.supports_routing_replay_capture()
+    captured: list[torch.Tensor] = []
+    experts.set_capture_fn(lambda r: captured.append(r.clone()))
+
+    hidden_states = (
+        torch.randn(num_tokens, hidden_size, device=device, dtype=torch.bfloat16) * 0.1
+    )
+    hidden_states_q, hidden_states_scale = fp4_quantize(
+        hidden_states,
+        a_gs,
+        block_size,
+        sf_use_ue8m0=False,
+        is_sf_swizzled_layout=False,
+    )
+    # The vLLM apply() does the .view(fp8_e4m3fn).reshape itself, so leave
+    # ``hidden_states_scale`` in its native (uint8 packed) form.
+    router_logits = torch.rand(
+        num_tokens, num_experts, device=device, dtype=torch.float32
+    )
+    routing_bias = _make_dsv3_routing_bias(num_experts, device)
+
+    _ = _run_nvfp4_monolithic(
+        experts,
+        hidden_states_q=hidden_states_q,
+        hidden_states_scale=hidden_states_scale,
+        router_logits=router_logits,
+        num_experts=num_experts,
+        routing_bias=routing_bias,
+    )
+
+    assert len(captured) == 1
+    replay = captured[0]
+    assert replay.dtype == torch.int16
+    assert replay.shape == (num_tokens, top_k)
+    assert (replay >= 0).all(), f"got out-of-range values: {replay}"
+    assert (replay < num_experts).all(), f"got out-of-range values: {replay}"
+    for t in range(num_tokens):
+        unique = replay[t].unique()
+        assert unique.numel() == top_k, (
+            f"token {t}: expected {top_k} distinct experts, "
+            f"got {unique.numel()} ({replay[t].tolist()})"
+        )
@@ -66,7 +66,6 @@ def test_worker_apply_lora(qwen3_lora_files):
            runner_type="generate",
            max_num_batched_tokens=32,
            max_num_seqs=32,
-            max_num_partial_prefills=32,
        ),
        device_config=DeviceConfig(DEVICE_TYPE),
        cache_config=CacheConfig(
@@ -0,0 +1,306 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import ast
+from dataclasses import dataclass
+from types import SimpleNamespace
+from typing import Any, cast
+
+import pytest
+
+from vllm.model_executor.warmup.jit_warmup import (
+    VllmJitKernel,
+    WarmupIntRange,
+    get_ast_full_name,
+    zip_inputs,
+)
+
+
+def _next_power_of_2(value: int) -> int:
+    return 1 << max(0, value - 1).bit_length()
+
+
+def _round_up(value: int, *, multiple: int) -> int:
+    return ((value + multiple - 1) // multiple) * multiple
+
+
+def _config(
+    *,
+    bias: int = 0,
+    disabled: bool = False,
+    name: str = "base",
+    vectorized: bool = False,
+) -> SimpleNamespace:
+    return SimpleNamespace(
+        bias=bias,
+        disabled=disabled,
+        name=name,
+        vectorized=vectorized,
+    )
+
+
+class ToyKernel(VllmJitKernel["ToyKernel.CompileKey"]):
+    @dataclass(frozen=True)
+    class CompileKey:
+        block_size: int
+        work: int
+        vector_width: int
+        descriptor: tuple[object, ...]
+        enabled: bool
+
+    def dispatch(  # type: ignore[override]
+        self,
+        *,
+        tokens: int,
+        cfg: Any,
+        lanes: int = 1,
+        mode: str = "default",
+        debug: int = 0,
+    ) -> CompileKey:
+        block_size = _next_power_of_2(tokens)
+        work: int = block_size * lanes + cfg.bias
+        return self.CompileKey(
+            block_size=block_size,
+            work=work,
+            vector_width=4 if cfg.vectorized and block_size >= 4 else 1,
+            descriptor=(
+                cfg.name,
+                mode,
+                -block_size,
+                block_size % 3,
+                block_size**2,
+            ),
+            enabled=not cfg.disabled,
+        )
+
+    def get_warmup_keys(self, max_tokens: int, cfg: Any) -> list[CompileKey]:
+        return self._trace_dispatch(self.dispatch)(
+            tokens=WarmupIntRange(1, max_tokens + 1),
+            cfg=cfg,
+            # This argument is intentionally unused by dispatch expressions.
+            debug=WarmupIntRange(0, 100),
+        )
+
+    def compile(self, compile_key: CompileKey) -> None:
+        pass
+
+
+class RecordingToyKernel(ToyKernel):
+    def __init__(self) -> None:
+        self.compiled: list[ToyKernel.CompileKey] = []
+        super().__init__()
+
+    def compile(self, compile_key: ToyKernel.CompileKey) -> None:
+        self.compiled.append(compile_key)
+
+
+def test_trace_dispatch_expands_ranges_dedupes_and_ignores_unused_inputs() -> None:
+    cfg = _config()
+
+    assert ToyKernel().get_warmup_keys(5, cfg) == [
+        ToyKernel.CompileKey(1, 1, 1, ("base", "default", -1, 1, 1), True),
+        ToyKernel.CompileKey(2, 2, 1, ("base", "default", -2, 2, 4), True),
+        ToyKernel.CompileKey(4, 4, 1, ("base", "default", -4, 1, 16), True),
+        ToyKernel.CompileKey(8, 8, 1, ("base", "default", -8, 2, 64), True),
+    ]
+
+
+def test_compile_key_uses_defaults_locals_attributes_and_expressions() -> None:
+    cfg = _config(bias=3, disabled=True, name="cfg", vectorized=True)
+
+    assert ToyKernel().compile_key(
+        {
+            "tokens": 4,
+            "cfg": cfg,
+            "lanes": 2,
+        }
+    ) == ToyKernel.CompileKey(
+        block_size=4,
+        work=11,
+        vector_width=4,
+        descriptor=("cfg", "default", -4, 1, 16),
+        enabled=False,
+    )
+
+
+def test_trace_dispatch_combines_zipped_rows_with_independent_values() -> None:
+    cfg = _config(vectorized=True)
+
+    keys = ToyKernel()._trace_dispatch(ToyKernel().dispatch)(
+        zip_inputs(
+            dict(tokens=1, mode="small"),
+            dict(tokens=4, mode="wide"),
+        ),
+        cfg=cfg,
+        lanes=(1, 2),
+    )
+
+    assert keys == [
+        ToyKernel.CompileKey(1, 1, 1, ("base", "small", -1, 1, 1), True),
+        ToyKernel.CompileKey(1, 2, 1, ("base", "small", -1, 1, 1), True),
+        ToyKernel.CompileKey(4, 4, 4, ("base", "wide", -4, 1, 16), True),
+        ToyKernel.CompileKey(4, 8, 4, ("base", "wide", -4, 1, 16), True),
+    ]
+
+
+def test_zip_inputs_validates_input_rows() -> None:
+    with pytest.raises(ValueError, match="requires at least one"):
+        zip_inputs()
+    with pytest.raises(ValueError, match="rows must be mappings"):
+        zip_inputs(cast(Any, ("tokens", 1)))
+    with pytest.raises(ValueError, match="at least one dispatch input name"):
+        zip_inputs({})
+    with pytest.raises(ValueError, match="dispatch input names must be strings"):
+        zip_inputs(cast(Any, {1: 2}))
+    with pytest.raises(ValueError, match="same dispatch input names"):
+        zip_inputs({"tokens": 1}, {"mode": "small"})
+
+
+def test_trace_dispatch_rejects_bad_positional_groups_and_duplicates() -> None:
+    kernel = ToyKernel()
+
+    with pytest.raises(TypeError, match="zip_inputs"):
+        kernel._trace_dispatch(kernel.dispatch)(
+            cast(Any, {"tokens": 1}),
+            cfg=_config(),
+        )
+
+    with pytest.raises(ValueError, match="specified more than once"):
+        kernel._trace_dispatch(kernel.dispatch)(
+            zip_inputs(dict(tokens=1, mode="small")),
+            tokens=2,
+            cfg=_config(),
+        )
+
+
+def test_helper_calls_support_keywords_and_reject_star_kwargs() -> None:
+    class HelperKernel(VllmJitKernel["HelperKernel.CompileKey"]):
+        @dataclass(frozen=True)
+        class CompileKey:
+            value: int
+
+        def dispatch(  # type: ignore[override]
+            self,
+            *,
+            tokens: int,
+            block_size: int,
+        ) -> CompileKey:
+            return self.CompileKey(value=_round_up(tokens, multiple=block_size))
+
+        def get_warmup_keys(self) -> list[CompileKey]:
+            return []
+
+        def compile(self, compile_key: CompileKey) -> None:
+            pass
+
+    class StarKwargsKernel(VllmJitKernel["StarKwargsKernel.CompileKey"]):
+        @dataclass(frozen=True)
+        class CompileKey:
+            value: int
+
+        def dispatch(  # type: ignore[override]
+            self,
+            *,
+            tokens: int,
+            block_size: int,
+        ) -> CompileKey:
+            return self.CompileKey(value=_round_up(tokens, **{"multiple": block_size}))
+
+        def get_warmup_keys(self) -> list[CompileKey]:
+            return []
+
+        def compile(self, compile_key: CompileKey) -> None:
+            pass
+
+    assert HelperKernel().compile_key(
+        {
+            "tokens": 5,
+            "block_size": 4,
+        }
+    ) == HelperKernel.CompileKey(value=8)
+    with pytest.raises(ValueError, match=r"cannot use \*\*kwargs"):
+        StarKwargsKernel().compile_key({"tokens": 5, "block_size": 4})
+
+
+def test_dispatch_body_must_be_local_assignments_then_compile_key_return() -> None:
+    class BranchKernel(VllmJitKernel["BranchKernel.CompileKey"]):
+        @dataclass(frozen=True)
+        class CompileKey:
+            value: int
+
+        def dispatch(self, *, value: int) -> CompileKey:  # type: ignore[override]
+            if value > 0:
+                value = 1
+            return self.CompileKey(value=value)
+
+        def get_warmup_keys(self) -> list[CompileKey]:
+            return []
+
+        def compile(self, compile_key: CompileKey) -> None:
+            pass
+
+    class KwargsReturnKernel(VllmJitKernel["KwargsReturnKernel.CompileKey"]):
+        @dataclass(frozen=True)
+        class CompileKey:
+            value: int
+
+        def dispatch(self, *, value: int) -> CompileKey:  # type: ignore[override]
+            return self.CompileKey(**{"value": value})
+
+        def get_warmup_keys(self) -> list[CompileKey]:
+            return []
+
+        def compile(self, compile_key: CompileKey) -> None:
+            pass
+
+    with pytest.raises(ValueError, match="local assignments"):
+        BranchKernel()
+    with pytest.raises(ValueError, match=r"cannot use \*\*kwargs in CompileKey"):
+        KwargsReturnKernel()
+
+
+def test_dispatch_reports_unsupported_expression_with_context() -> None:
+    class UnsupportedKernel(VllmJitKernel["UnsupportedKernel.CompileKey"]):
+        @dataclass(frozen=True)
+        class CompileKey:
+            value: object
+
+        def dispatch(self, *, value: int) -> CompileKey:  # type: ignore[override]
+            return self.CompileKey(value={value})
+
+        def get_warmup_keys(self) -> list[CompileKey]:
+            return []
+
+        def compile(self, compile_key: CompileKey) -> None:
+            pass
+
+    with pytest.raises(ValueError) as exc_info:
+        UnsupportedKernel().compile_key({"value": 1})
+
+    message = str(exc_info.value)
+    assert "Unsupported dispatch expression" in message
+    assert "{value}" in message
+    assert "Supported dispatch expressions" in message
+
+
+def test_warmup_compiles_all_returned_keys_in_order() -> None:
+    kernel = RecordingToyKernel()
+    cfg = _config()
+
+    kernel.warmup(3, cfg)
+
+    assert kernel.compiled == [
+        ToyKernel.CompileKey(1, 1, 1, ("base", "default", -1, 1, 1), True),
+        ToyKernel.CompileKey(2, 2, 1, ("base", "default", -2, 2, 4), True),
+        ToyKernel.CompileKey(4, 4, 1, ("base", "default", -4, 1, 16), True),
+    ]
+
+
+def test_get_ast_full_name_handles_names_attributes_and_other_nodes() -> None:
+    dotted_expr = ast.parse("foo.bar.baz").body[0]
+    call_expr = ast.parse("foo()").body[0]
+    assert isinstance(dotted_expr, ast.Expr)
+    assert isinstance(call_expr, ast.Expr)
+
+    assert get_ast_full_name(dotted_expr.value) == "foo.bar.baz"
+    assert get_ast_full_name(call_expr.value) is None
@@ -66,6 +66,12 @@ def _make_router(eplb_state: EplbLayerState | None = None) -> DummyRouter:
    )


+def _make_modular_routed_experts():
+    return types.SimpleNamespace(
+        quant_method=types.SimpleNamespace(is_monolithic=False),
+    )
+
+
 def test_base_router_capture_pre_eplb_mapping():
    router = _make_router()
    captured = []
@@ -122,6 +128,8 @@ def test_gpu_model_runner_binds_router_capture(monkeypatch):
        def __init__(self):
            self.layer_id = 7
            self.router = _make_router()
+            self.routed_experts = _make_modular_routed_experts()
+            self._quant_method = self.routed_experts.quant_method

    class DummyCapturer:
        def __init__(self):
@@ -160,6 +168,8 @@ def test_gpu_model_runner_binding_stage(monkeypatch):
        def __init__(self):
            self.layer_id = 11
            self.router = _make_router()
+            self.routed_experts = _make_modular_routed_experts()
+            self._quant_method = self.routed_experts.quant_method

    class DummyCapturer:
        def __init__(self):
@@ -197,6 +207,8 @@ def test_gpu_model_runner_does_not_bind_draft_router_capture(monkeypatch):
        def __init__(self, layer_id):
            self.layer_id = layer_id
            self.router = _make_router()
+            self.routed_experts = _make_modular_routed_experts()
+            self._quant_method = self.routed_experts.quant_method

    target_module = DummyFusedMoE(layer_id=7)
    draft_module = DummyFusedMoE(layer_id=0)
@@ -222,6 +234,49 @@ def test_gpu_model_runner_does_not_bind_draft_router_capture(monkeypatch):
    assert draft_module.router.capture_fn is None


+def test_gpu_model_runner_rejects_monolithic_without_replay_support(monkeypatch):
+    from vllm.v1.worker import gpu_model_runner as gmr
+
+    class DummyFusedMoE:
+        def __init__(self):
+            self.layer_id = 3
+            self.router = _make_router()
+            # Use a concrete monolithic expert and override its capability
+            # instead of instantiating the abstract base class directly.
+            from vllm.model_executor.layers.fused_moe.experts.cpu_moe import (
+                CPUExpertsFp8,
+            )
+
+            fused_experts = CPUExpertsFp8.__new__(CPUExpertsFp8)
+            self.routed_experts = types.SimpleNamespace(
+                quant_method=types.SimpleNamespace(
+                    is_monolithic=True,
+                    moe_kernel=types.SimpleNamespace(
+                        impl=types.SimpleNamespace(fused_experts=fused_experts)
+                    ),
+                )
+            )
+            self._quant_method = self.routed_experts.quant_method
+            self._quant_method.moe_kernel.impl.fused_experts = fused_experts
+            fused_experts.supports_routing_replay_capture = lambda: False
+
+    class DummyCapturer:
+        def capture(self, layer_id, topk_ids):
+            pass
+
+    dummy_module = DummyFusedMoE()
+    import vllm.model_executor.layers.fused_moe.layer as fused_moe_layer
+
+    monkeypatch.setattr(fused_moe_layer, "MoERunner", DummyFusedMoE)
+
+    dummy_self = types.SimpleNamespace(
+        model=types.SimpleNamespace(modules=lambda: [dummy_module])
+    )
+
+    with pytest.raises(ValueError, match="monolithic MoE kernel"):
+        gmr.GPUModelRunner._bind_routed_experts_capturer(dummy_self, DummyCapturer())
+
+
 def test_routed_experts_capturer_single_dp_no_metadata():
    """dp_metadata is None: capture writes the full topk_ids rows."""
    capturer = _capturer_with_buffer(dp_rank=0)
@@ -9,6 +9,7 @@ import torch.nn.functional as F
 from transformers import AutoModel

 from vllm.platforms import current_platform
+from vllm.utils.mem_constants import MiB_bytes

 from ....conftest import HfRunner
 from ....utils import VLLM_PATH
@@ -106,7 +107,12 @@ def test_prm_models(
    if current_platform.is_cpu():
        pytest.skip("CPU only supports V1")

-    with vllm_runner(model, max_model_len=1024, dtype=dtype) as vllm_model:
+    with vllm_runner(
+        model,
+        max_model_len=1024,
+        dtype=dtype,
+        kv_cache_memory_bytes=64 * MiB_bytes,
+    ) as vllm_model:
        vllm_outputs = vllm_model.token_classify(math_step_prompts)

    with hf_runner(model, dtype=dtype, auto_cls=AutoModel) as hf_model:
@@ -145,7 +151,12 @@ def test_prm_models_with_golden_outputs(
    if not FIXTURE_REWARD_RESULT.get(model):
        pytest.skip(f"No available golden outputs for {model}.")

-    with vllm_runner(model, max_model_len=1024, dtype=dtype) as vllm_model:
+    with vllm_runner(
+        model,
+        max_model_len=1024,
+        dtype=dtype,
+        kv_cache_memory_bytes=64 * MiB_bytes,
+    ) as vllm_model:
        vllm_outputs = vllm_model.token_classify(math_step_prompts)

    golden_outputs = load_reward_outputs(FIXTURE_REWARD_RESULT[model])
@@ -70,7 +70,19 @@ def _assert_video_outputs(processor, processed) -> None:
    merge_size = processor.info.get_hf_config().vision_config.spatial_merge_size
    expected_tokens = int(grid_thw.prod()) // merge_size**2
    video_token_id = processor.info.get_hf_config().video_token_id
-    assert processed["prompt_token_ids"].count(video_token_id) == expected_tokens
+    prompt_token_ids = processed["prompt_token_ids"]
+    assert prompt_token_ids.count(video_token_id) == expected_tokens
+
+    hf_processor = processor.info.get_hf_processor()
+    expected_frame_wrappers = int(grid_thw[:, 0].sum())
+    assert (
+        prompt_token_ids.count(hf_processor.vision_start_token_id)
+        == expected_frame_wrappers
+    )
+    assert (
+        prompt_token_ids.count(hf_processor.vision_end_token_id)
+        == expected_frame_wrappers
+    )


@pytest.mark.parametrize("num_images", [1, 2])
@@ -74,6 +74,44 @@ def test_cosmos3_new_checkpoint_weights_mapper():
    )


+def test_cosmos3_modelopt_quantizer_weights_mapper():
+    """ModelOpt/Diffusers FP8 checkpoints ship native fake-quant buffers
+    (``*_quantizer._amax`` / ``._scale``) alongside the vLLM-consumable
+    ``weight_scale`` / ``input_scale`` sidecars. vLLM must drop the former
+    (it has no parameter for them) while keeping the latter."""
+    from vllm.model_executor.models.cosmos3 import Cosmos3ForConditionalGeneration
+
+    mapper = Cosmos3ForConditionalGeneration.hf_to_vllm_mapper
+
+    # Native ModelOpt quantizer buffers are dropped.
+    assert (
+        mapper.apply_list(
+            [
+                "layers.0.self_attn.to_q.input_quantizer._amax",
+                "layers.0.self_attn.to_q.weight_quantizer._amax",
+                "layers.0.self_attn.to_q.weight_quantizer._scale",
+                "layers.0.mlp.down_proj.output_quantizer._amax",
+            ]
+        )
+        == []
+    )
+
+    # The FP8 scale sidecars vLLM actually consumes are kept and remapped.
+    assert mapper.apply_list(
+        [
+            "layers.0.self_attn.to_q.weight",
+            "layers.0.self_attn.to_q.weight_scale",
+            "layers.0.self_attn.to_q.input_scale",
+            "layers.0.mlp.down_proj.input_scale",
+        ]
+    ) == [
+        "language_model.model.layers.0.self_attn.q_proj.weight",
+        "language_model.model.layers.0.self_attn.q_proj.weight_scale",
+        "language_model.model.layers.0.self_attn.q_proj.input_scale",
+        "language_model.model.layers.0.mlp.down_proj.input_scale",
+    ]
+
+
 def test_cosmos3_edge_checkpoint_weights_mapper():
    from vllm.model_executor.models.cosmos3_edge import (
        Cosmos3EdgeForConditionalGeneration,
@@ -132,6 +170,7 @@ def test_cosmos3_edge_checkpoint_weights_mapper():
                "layers.0.self_attn.to_add_out.weight",
                "layers.0.self_attn.norm_added_q.weight",
                "layers.0.self_attn.norm_added_k.weight",
+                "layers.0.self_attn.k_norm_und_for_gen.weight",
                "layers.0.self_attn.q_proj_moe_gen.weight",
                "layers.0.mlp_moe_gen.up_proj.weight",
                "norm_moe_gen.weight",
@@ -9,8 +9,8 @@ Note: these tests will only pass on L4 GPU.
 import pytest

 from tests.quantization.utils import is_quant_method_supported
+from vllm.v1.attention.backends.fa_utils import flash_attn_supports_kv_cache_dtype
 from vllm.platforms import current_platform
-from vllm.v1.attention.backends.fa_utils import get_flash_attn_version
 from ..utils import check_logprobs_close


@@ -70,13 +70,7 @@ def test_models(
    if kv_cache_dtype == "fp8_e5m2" and current_platform.is_cuda():
        pytest.skip(f"{kv_cache_dtype} is not supported by FLASH_ATTN on CUDA.")

-    if not (
-        current_platform.is_xpu()
-        or (
-            get_flash_attn_version() == 3
-            and current_platform.is_device_capability_family(90)
-        )
-    ):
+    if not flash_attn_supports_kv_cache_dtype(kv_cache_dtype):
        pytest.skip(
            f"{kv_cache_dtype} is not supported on this GPU type with {backend} attention."
        )
@@ -6,6 +6,7 @@ import mimetypes
 import os
 import shutil
 import time
+from io import BytesIO
 from tempfile import NamedTemporaryFile, TemporaryDirectory

 import aiohttp
@@ -111,6 +112,34 @@ async def test_fetch_image_base64(
        assert _image_equals(data_image_sync, data_image_async)


+@pytest.mark.asyncio
+async def test_fetch_image_keep_original_mode():
+    """media_io_kwargs can disable the default RGB conversion."""
+    # RGBA image: opaque black pixel on a fully transparent background
+    rgba_image = Image.new("RGBA", (4, 4), (0, 0, 0, 0))
+    rgba_image.putpixel((2, 2), (0, 0, 0, 255))
+    buffer = BytesIO()
+    rgba_image.save(buffer, "PNG")
+    data_url = (
+        f"data:image/png;base64,{base64.b64encode(buffer.getvalue()).decode('utf-8')}"
+    )
+
+    # Default behavior: RGBA is composited onto a white background
+    default_image = MediaConnector().fetch_image(data_url)
+    assert default_image.mode == "RGB"
+    assert default_image.getpixel((0, 0)) == (255, 255, 255)
+    assert default_image.getpixel((2, 2)) == (0, 0, 0)
+
+    # image_mode=None via media_io_kwargs: original mode is preserved
+    connector = MediaConnector(media_io_kwargs={"image": {"image_mode": None}})
+    image_sync = connector.fetch_image(data_url)
+    image_async = await connector.fetch_image_async(data_url)
+    for image in (image_sync, image_async):
+        assert image.mode == "RGBA"
+        assert image.getpixel((0, 0)) == (0, 0, 0, 0)
+        assert image.getpixel((2, 2)) == (0, 0, 0, 255)
+
+
@pytest.mark.asyncio
@pytest.mark.parametrize("image_url", TEST_IMAGE_ASSETS, indirect=True)
 async def test_fetch_image_local_files(image_url: str):
@@ -80,6 +80,29 @@ def test_image_media_io_rgba_custom_background(tmp_path):
    assert green_numpy[0][0][2] == 0  # B


+def test_image_media_io_no_mode_conversion(tmp_path):
+    """image_mode=None skips conversion and preserves the original mode."""
+    # RGBA image: opaque black pixel on a fully transparent background
+    rgba_image = Image.new("RGBA", (10, 10), (0, 0, 0, 0))
+    rgba_image.putpixel((5, 5), (0, 0, 0, 255))
+    test_image_path = tmp_path / "test_rgba.png"
+    rgba_image.save(test_image_path)
+
+    # Default behavior: RGBA is composited onto a white background
+    image_io_default = ImageMediaIO()
+    converted_default = image_io_default.load_file(test_image_path)
+    assert converted_default.media.mode == "RGB"
+    assert converted_default.media.getpixel((0, 0)) == (255, 255, 255)
+    assert converted_default.media.getpixel((5, 5)) == (0, 0, 0)
+
+    # image_mode=None: original mode and alpha channel are preserved
+    image_io_keep = ImageMediaIO(image_mode=None)
+    converted_keep = image_io_keep.load_file(test_image_path)
+    assert converted_keep.media.mode == "RGBA"
+    assert converted_keep.media.getpixel((0, 0)) == (0, 0, 0, 0)
+    assert converted_keep.media.getpixel((5, 5)) == (0, 0, 0, 255)
+
+
 def test_image_media_io_rgba_background_color_validation():
    """Test that invalid rgba_background_color values are properly rejected."""

@@ -304,6 +304,13 @@ def test_pynvvideocodec_decoder_slot_retains_simple_decoder():
 # ============================================================================


+def test_cosmos3_edge_uses_qwen3_vl_video_backend():
+    backend = get_video_loader_backend_for_processor("Cosmos3EdgeVideoProcessor")
+
+    assert backend == "qwen3_vl"
+    assert isinstance(VIDEO_LOADER_REGISTRY.load(backend), Qwen3VLVideoBackend)
+
+
@pytest.mark.parametrize(
    "model_repo, expected_loader_cls, hf_sample_kwargs",
    [
@@ -48,12 +48,12 @@ def _check_dense_embedding(data, index=0):
 def _check_sparse_embedding(data, check_tokens=False):
    expected_weights = [
        {"token_id": 32, "weight": 0.0552978515625, "token": "?"},
-        {"token_id": 70, "weight": 0.09808349609375, "token": "the"},
-        {"token_id": 83, "weight": 0.08154296875, "token": "is"},
-        {"token_id": 111, "weight": 0.11810302734375, "token": "of"},
-        {"token_id": 4865, "weight": 0.1171875, "token": "What"},
-        {"token_id": 9942, "weight": 0.292236328125, "token": "France"},
-        {"token_id": 10323, "weight": 0.2802734375, "token": "capital"},
+        {"token_id": 70, "weight": 0.09808349609375, "token": " the"},
+        {"token_id": 83, "weight": 0.08154296875, "token": " is"},
+        {"token_id": 111, "weight": 0.11810302734375, "token": " of"},
+        {"token_id": 4865, "weight": 0.1171875, "token": " What"},
+        {"token_id": 9942, "weight": 0.292236328125, "token": " France"},
+        {"token_id": 10323, "weight": 0.2802734375, "token": " capital"},
    ]
    expected_embed = {x["token_id"]: x for x in expected_weights}

@@ -93,9 +93,6 @@ def test_online_quantization(
    use_rocm_aiter: bool,
    monkeypatch,
 ) -> None:
-    if kv_cache_dtype == "fp8" and current_platform.is_device_capability_family(90):
-        pytest.skip("FA3 currently rejects FP8 KV cache output dtype on SM90")
-
    if use_rocm_aiter:
        monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")

@@ -105,9 +102,15 @@ def test_online_quantization(
    if force_marlin:
        monkeypatch.setenv("VLLM_TEST_FORCE_FP8_MARLIN", "1")

+    model_dtype = "auto"
+    if kv_cache_dtype == "fp8" and current_platform.is_device_capability_family(90):
+        # FA3 requires BF16 output when the query input is FP8.
+        model_dtype = "bfloat16"
+
    with vllm_runner(
        "facebook/opt-125m",
        quantization="fp8",
+        dtype=model_dtype,
        enforce_eager=True,
        kv_cache_dtype=kv_cache_dtype,
    ) as llm:
@@ -165,6 +165,38 @@ def test_modelopt_mixed_precision_does_not_quantize_unlisted_fused_sibling():
    assert config._resolve_quant_algo("model.layers.0.linear_attn.in_proj_ba") is None


+def test_modelopt_mixed_precision_composes_gemma4_mappers():
+    from vllm.model_executor.models.gemma4 import Gemma4ForCausalLM
+    from vllm.model_executor.models.gemma4_mm import (
+        Gemma4ForConditionalGeneration,
+    )
+
+    config = _mixed_precision_config(
+        {
+            "model.language_model.layers.0.experts": {
+                "quant_algo": "NVFP4",
+                "group_size": 16,
+            },
+            "model.language_model.layers.1.moe.experts.gate_up_proj": {
+                "quant_algo": "NVFP4",
+                "group_size": 16,
+            },
+        }
+    )
+
+    config.apply_vllm_mapper(
+        Gemma4ForConditionalGeneration.hf_to_vllm_mapper.get_unstacked_mapper()
+    )
+    config.apply_vllm_mapper(Gemma4ForCausalLM.hf_to_vllm_mapper.get_unstacked_mapper())
+
+    expected_prefix = "language_model.model.layers.0.moe.experts"
+    assert set(config.quantized_layers) == {
+        expected_prefix,
+        "language_model.model.layers.1.moe.gate_up_proj",
+    }
+    assert config._resolve_quant_algo(expected_prefix) == "NVFP4"
+
+
 def test_modelopt_mixed_precision_infers_fused_gate_up_projection():
    from vllm.model_executor.layers.linear import LinearBase

@@ -1,141 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import pytest
-import torch
-
-from vllm.utils.extensible_tensor import ExtensibleTensor
-
-pytestmark = pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA")
-
-
-def test_extensible_tensor_grows_without_moving() -> None:
-    buffer = ExtensibleTensor(4096, device="cuda")
-    try:
-        base_ptr = buffer.base_ptr
-        first_view = buffer.resize_(1024)
-        assert first_view.data_ptr() == base_ptr
-        first_view.fill_(7)
-
-        second_view = buffer.resize_(2048)
-        assert second_view.data_ptr() == base_ptr
-        assert torch.equal(second_view[:1024], torch.full_like(second_view[:1024], 7))
-
-        second_view[1024:].fill_(3)
-        assert torch.equal(buffer.tensor, second_view)
-
-        full_view = buffer.full_view()
-        assert full_view.data_ptr() == base_ptr
-        assert full_view.numel() == 4096
-    finally:
-        buffer.free()
-
-
-def test_extensible_tensor_rejects_shrink_and_overflow() -> None:
-    buffer = ExtensibleTensor(1024, device="cuda")
-    try:
-        buffer.resize_(512)
-        with pytest.raises(ValueError, match="grow-only"):
-            buffer.resize_(256)
-        with pytest.raises(ValueError, match="exceeds the segment capacity"):
-            buffer.resize_(1025)
-    finally:
-        buffer.free()
-
-
-def test_segments_grow_in_lockstep_and_zero_new() -> None:
-    """Each segment's committed prefix grows in lockstep.
-
-    Data written to a segment's committed prefix survives a grow; the newly
-    committed range of each segment is zeroed with `zero_new=True` while old
-    bytes are preserved.
-    """
-    et = ExtensibleTensor(max_num_bytes=8192, device="cuda", num_segments=2)
-    try:
-        assert et.num_segments == 2
-        assert et.segment_capacity_bytes == 4096
-
-        et.resize_per_segment_(256, zero_new=True)
-        assert et.bytes_per_segment == 256
-        assert et.num_bytes == 512
-        fv = et.full_view()
-        assert fv.shape == (8192,)
-        # Committed prefixes start zeroed.
-        assert torch.count_nonzero(fv[:256]) == 0
-        assert torch.count_nonzero(fv[4096 : 4096 + 256]) == 0
-
-        pattern_a = torch.arange(256, device="cuda", dtype=torch.uint8)
-        pattern_b = 255 - pattern_a
-        fv[:256].copy_(pattern_a)
-        fv[4096 : 4096 + 256].copy_(pattern_b)
-
-        et.resize_per_segment_(1024, zero_new=True)
-        fv2 = et.full_view()
-        assert fv2.data_ptr() == fv.data_ptr()
-        # Old bytes of both segments preserved; freshly committed ranges zeroed.
-        assert torch.equal(fv2[:256], pattern_a)
-        assert torch.equal(fv2[4096 : 4096 + 256], pattern_b)
-        assert torch.count_nonzero(fv2[256:1024]) == 0
-        assert torch.count_nonzero(fv2[4096 + 256 : 4096 + 1024]) == 0
-    finally:
-        et.free()
-
-
-def test_segments_at_granularity_scale() -> None:
-    """Segments spanning multiple mapping granules commit correctly.
-
-    Uses a segment capacity that is not a multiple of the allocation
-    granularity, so a granule straddles the segment boundary and is shared by
-    the first commit of one segment and a later commit of the other -- it must
-    be mapped exactly once.
-    """
-    probe = ExtensibleTensor(max_num_bytes=1, device="cuda")
-    granularity = probe.capacity_bytes
-    probe.free()
-    # Two segments of 1.5 granules each; the middle granule straddles the
-    # boundary.
-    max_num_bytes = 3 * granularity
-    et = ExtensibleTensor(max_num_bytes=max_num_bytes, device="cuda", num_segments=2)
-    try:
-        seg = et.segment_capacity_bytes
-        assert seg == max_num_bytes // 2
-
-        step = granularity // 2
-        et.resize_per_segment_(step, zero_new=True)
-        fv = et.full_view()
-        fv[:step].fill_(1)
-        fv[seg : seg + step].fill_(2)
-
-        # Grow to the full segment capacity: previously mapped granules
-        # (including the boundary-straddling one) are reused, new ones are
-        # committed and zeroed.
-        et.resize_per_segment_(seg, zero_new=True)
-        fv2 = et.full_view()
-        assert torch.all(fv2[:step] == 1)
-        assert torch.all(fv2[seg : seg + step] == 2)
-        assert torch.count_nonzero(fv2[step:seg]) == 0
-        assert torch.count_nonzero(fv2[seg + step :]) == 0
-    finally:
-        et.free()
-
-
-def test_multi_segment_invalid_usage_raises() -> None:
-    """Prefix-view APIs and invalid segment configs raise for multi-segment
-    buffers."""
-    with pytest.raises(ValueError):
-        ExtensibleTensor(max_num_bytes=100, device="cuda", num_segments=3)
-
-    et = ExtensibleTensor(max_num_bytes=8192, device="cuda", num_segments=2)
-    try:
-        with pytest.raises(ValueError):
-            _ = et.tensor
-        with pytest.raises(ValueError):
-            et.resize_(256)
-
-        et.resize_per_segment_(256)
-        with pytest.raises(ValueError):
-            et.resize_per_segment_(128)  # shrink
-        with pytest.raises(ValueError):
-            et.resize_per_segment_(et.segment_capacity_bytes + 1)  # over capacity
-    finally:
-        et.free()
@@ -67,6 +67,23 @@ def test_memory_profiling():
    non_torch_ratio = result.non_torch_increase / (256 * 1024 * 1024)  # noqa
    assert abs(non_torch_ratio - 1) <= 0.05
    assert result.torch_peak_increase == 1024 * 1024 * 1024
+
+    expected_total_consumed = (256 + 512) * 1024 * 1024
+    total_consumed_ratio = result.total_consumed / expected_total_consumed
+    assert abs(total_consumed_ratio - 1) <= 0.05, (
+        f"total_consumed={result.total_consumed}, "
+        f"expected={expected_total_consumed}, "
+        f"ratio={total_consumed_ratio}"
+    )
+
+    expected_non_kv = expected_total_consumed + 1024 * 1024 * 1024
+    non_kv_ratio = result.non_kv_cache_memory / expected_non_kv
+    assert abs(non_kv_ratio - 1) <= 0.05, (
+        f"non_kv_cache_memory={result.non_kv_cache_memory}, "
+        f"expected={expected_non_kv}, "
+        f"ratio={non_kv_ratio}"
+    )
+
    del weights
    lib.cudaFree(handle1)
    lib.cudaFree(handle2)
@@ -21,6 +21,7 @@ from vllm.platforms import current_platform
 from vllm.utils.math_utils import cdiv
 from vllm.utils.torch_utils import (
    STR_DTYPE_TO_TORCH_DTYPE,
+    is_quantized_kv_cache,
    is_torch_equal_or_newer,
    set_random_seed,
 )
@@ -45,6 +46,11 @@ BACKENDS_TO_TEST = [

 DEVICE_TYPE = current_platform.device_type

+FP8_KV_CACHE_DTYPES = {
+    "fp8": torch.float8_e4m3fn,
+    "fp8_e4m3": torch.float8_e4m3fn,
+}
+
 # Remove flashinfer from the list if it's not available
 try:
    import flashinfer  # noqa: F401
@@ -110,6 +116,7 @@ def create_and_prepopulate_kv_cache(
    num_blocks: int,
    common_attn_metadata: CommonAttentionMetadata,
    randomize_blocks: bool = True,
+    kv_cache_dtype: str = "auto",
 ) -> torch.Tensor:
    """Create and prepopulate a KV cache with context data.

@@ -140,8 +147,18 @@ def create_and_prepopulate_kv_cache(
    block_table = common_attn_metadata.block_table_tensor
    slot_mapping = common_attn_metadata.slot_mapping

+    # For an fp8 kv cache, store the cache in the fp8 dtype so that assigning
+    # the higher-precision context tensors quantizes them, mirroring runtime.
+    fp8_kv_cache = is_quantized_kv_cache(kv_cache_dtype)
+    storage_dtype = FP8_KV_CACHE_DTYPES[kv_cache_dtype] if fp8_kv_cache else dtype
+
    kv_cache = torch.zeros(
-        num_blocks, block_size, num_kv_heads, 2 * head_size, dtype=dtype, device=device
+        num_blocks,
+        block_size,
+        num_kv_heads,
+        2 * head_size,
+        dtype=storage_dtype,
+        device=device,
    )
    kv_cache_flat = kv_cache.view(-1, num_kv_heads, 2 * head_size)

@@ -195,7 +212,12 @@ def create_and_prepopulate_kv_cache(
        ] * block_size + token_inter_block_offsets.to(device)

    # Transpose to logical (num_blocks, num_kv_heads, block_size, 2*hs)
-    return kv_cache.transpose(1, 2).contiguous()
+    kv_cache = kv_cache.transpose(1, 2).contiguous()
+
+    if fp8_kv_cache:
+        kv_cache = kv_cache.view(torch.uint8)
+
+    return kv_cache


 class MockAttentionLayer:
@@ -224,6 +246,7 @@ def run_attention_backend(
    kv_cache: torch.Tensor,
    attn_type: AttentionType = AttentionType.DECODER,
    sliding_window: int | None = None,
+    kv_cache_dtype: str = "auto",
 ) -> torch.Tensor:
    """Run attention computation using the specified backend's AttentionImpl."""

@@ -291,13 +314,16 @@ def run_attention_backend(
        alibi_slopes=None,
        sliding_window=sliding_window,
        attn_type=attn_type,
-        kv_cache_dtype="auto",
+        kv_cache_dtype=kv_cache_dtype,
    )

    # Create mock layer and output buffer
    mock_layer = MockAttentionLayer(device)
    output = torch.empty_like(query)

+    if is_quantized_kv_cache(kv_cache_dtype) and impl.supports_quant_query_input:
+        query = query.to(current_platform.fp8_dtype())
+
    # Run forward pass
    # NOTE: The query, key, and value are already shaped correctly
    # in the calling test function.
@@ -324,6 +350,7 @@ def _test_backend_correctness(
    atol: float = 1e-2,
    rtol: float = 1e-2,
    tensor_parallel_size: int = 1,
+    kv_cache_dtype: str = "auto",
 ):
    """
    Test that all backends produce similar outputs to a reference implementation
@@ -372,6 +399,7 @@ def _test_backend_correctness(
        num_gpu_blocks=8192,
        hf_config_override=hf_config_override,
    )
+    vllm_config.cache_config.cache_dtype = kv_cache_dtype
    device = torch.device(f"{DEVICE_TYPE}:0")

    kv_cache_spec = create_standard_kv_cache_spec(vllm_config, attn_type)
@@ -392,6 +420,13 @@ def _test_backend_correctness(
    block_size = vllm_config.cache_config.block_size
    scale = 1.0 / (head_size**0.5)

+    fp8_kv_cache = is_quantized_kv_cache(kv_cache_dtype)
+    if fp8_kv_cache:
+        query_fp8_dtype = current_platform.fp8_dtype()
+        kv_fp8_dtype = FP8_KV_CACHE_DTYPES[kv_cache_dtype]
+        atol = max(atol, 6e-2)
+        rtol = max(rtol, 1e-1)
+
    # 2. Generate data and compute SDPA reference output
    all_q_vllm, all_k_vllm, all_v_vllm = [], [], []
    all_sdpa_outputs = []
@@ -407,10 +442,17 @@ def _test_backend_correctness(
        k_full = torch.randn(s_len, num_kv_heads, head_size, dtype=dtype, device=device)
        v_full = torch.randn(s_len, num_kv_heads, head_size, dtype=dtype, device=device)

+        if fp8_kv_cache:
+            q_ref = q.to(query_fp8_dtype).to(dtype)
+            k_ref = k_full.to(kv_fp8_dtype).to(dtype)
+            v_ref = v_full.to(kv_fp8_dtype).to(dtype)
+        else:
+            q_ref, k_ref, v_ref = q, k_full, v_full
+
        # SDPA expects (N, H, L, D), so unsqueeze batch and permute
-        q_sdpa_in = q.unsqueeze(0).transpose(1, 2)
-        k_sdpa_in = k_full.unsqueeze(0).transpose(1, 2)
-        v_sdpa_in = v_full.unsqueeze(0).transpose(1, 2)
+        q_sdpa_in = q_ref.unsqueeze(0).transpose(1, 2)
+        k_sdpa_in = k_ref.unsqueeze(0).transpose(1, 2)
+        v_sdpa_in = v_ref.unsqueeze(0).transpose(1, 2)

        if num_q_heads != num_kv_heads:
            assert num_q_heads % num_kv_heads == 0, (
@@ -471,6 +513,7 @@ def _test_backend_correctness(
        num_blocks=vllm_config.cache_config.num_gpu_blocks or 1000,
        common_attn_metadata=common_attn_metadata,
        randomize_blocks=True,
+        kv_cache_dtype=kv_cache_dtype,
    )

    # 4. Run vLLM backends and compare
@@ -488,6 +531,12 @@ def _test_backend_correctness(
        else:
            backend_cls = None

+        if is_quantized_kv_cache(kv_cache_dtype) and (
+            backend_cls is None
+            or not backend_cls.supports_kv_cache_dtype(kv_cache_dtype)
+        ):
+            continue
+
        if backend_name == AttentionBackendEnum.FLASHINFER:
            set_kv_cache_layout("HND")
            reset_kv_cache_layout = True
@@ -521,6 +570,7 @@ def _test_backend_correctness(
                kv_cache_for_backend,
                sliding_window=sliding_window,
                attn_type=attn_type,
+                kv_cache_dtype=kv_cache_dtype,
            )
        finally:
            if reset_kv_cache_layout:
@@ -570,8 +620,13 @@ def _test_backend_correctness(
 )
@pytest.mark.parametrize("model", ["meta-llama/Meta-Llama-3-8B"])
@pytest.mark.parametrize("tensor_parallel_size", [1, 2, 4])
+@pytest.mark.parametrize("kv_cache_dtype", ["auto", "fp8", "fp8_e4m3"])
 def test_causal_backend_correctness(
-    default_vllm_config, batch_spec_name: str, model: str, tensor_parallel_size: int
+    default_vllm_config,
+    batch_spec_name: str,
+    model: str,
+    tensor_parallel_size: int,
+    kv_cache_dtype: str,
 ):
    """Test backend's correctness with causal attention."""

@@ -612,6 +667,7 @@ def test_causal_backend_correctness(
        SMALL_BLOCK_BACKENDS,
        causal_mask_mod,
        tensor_parallel_size=tensor_parallel_size,
+        kv_cache_dtype=kv_cache_dtype,
    )

    # Fast FlexAttention needs to run with block_size=128
@@ -623,6 +679,7 @@ def test_causal_backend_correctness(
            causal_mask_mod,
            block_size=128,
            tensor_parallel_size=tensor_parallel_size,
+            kv_cache_dtype=kv_cache_dtype,
        )


@@ -119,6 +119,7 @@ def test_mla_post_load_preserves_runtime_weight_addresses(monkeypatch):
    layer.kv_b_proj.quant_method = None
    layer.is_aiter_triton_fp4_bmm_enabled = False
    layer.is_aiter_triton_fp8_bmm_enabled = False
+    layer.dcp_q_replicate = False
    layer.quant_config = None
    layer.layer_name = "test"

@@ -33,6 +33,7 @@ from vllm.v1.kv_cache_interface import (
    EncoderOnlyAttentionSpec,
    FullAttentionSpec,
    MambaSpec,
+    get_kv_quant_mode,
 )


@@ -178,6 +179,7 @@ def create_standard_kv_cache_spec(
        head_size=vllm_config.model_config.get_head_size(),
        dtype=vllm_config.model_config.dtype,
        sliding_window=vllm_config.model_config.get_sliding_window(),
+        kv_quant_mode=get_kv_quant_mode(vllm_config.cache_config.cache_dtype),
    )


@@ -149,30 +149,6 @@ def test_has_cache_restores_from_freeable():
    assert manager.num_freeable_slots == 6


-def test_make_profiling_reservation():
-    assert (
-        EncoderCacheManager.make_profiling_reservation(
-            cache_size=0,
-            embed_size=8,
-            dtype=torch.float16,
-            device="cpu",
-        )
-        is None
-    )
-
-    reservation = EncoderCacheManager.make_profiling_reservation(
-        cache_size=7,
-        embed_size=8,
-        dtype=torch.float16,
-        device="cpu",
-    )
-
-    assert reservation is not None
-    assert reservation.shape == (7, 8)
-    assert reservation.dtype == torch.float16
-    assert reservation.device.type == "cpu"
-
-
 def test_get_freed_mm_hashes_clears_freed_list():
    manager = EncoderCacheManager(cache_size=10)
    req1 = MockRequest("reqA", ["a"], [5])
@@ -1,5 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import copy
 import hashlib
 import importlib
 from collections.abc import Callable
@@ -1477,6 +1478,128 @@ def test_get_max_concurrency_for_kv_cache_config():
    assert num_tokens == max_concurrency_hybrid_model * max_model_len
    assert max_concurrency == max_concurrency_hybrid_model

+    # Unequal group sizes in the standard layout: each group's pages cost
+    # whole pool blocks, so a request needs 1024 + 129 = 1153 blocks — the
+    # same as the equal-hybrid case above, regardless of the second group
+    # holding only 2 layers.
+    kv_cache_config_unequal_groups = KVCacheConfig(
+        num_blocks=1153 * 3,
+        kv_cache_tensors=[],
+        kv_cache_groups=[
+            KVCacheGroupSpec([f"layer_{i}" for i in range(32)], full_attention_spec),
+            KVCacheGroupSpec(["layer_32", "layer_33"], sliding_window_spec),
+        ],
+    )
+    assert (
+        get_max_concurrency_for_kv_cache_config(
+            vllm_config, kv_cache_config_unequal_groups
+        )
+        == 3
+    )
+
+    # UniformTypeKVCacheSpecs group (worker config shape): the aggregated
+    # spec's memory/page ratio equals a single layer's page count, so the
+    # group needs 1024 blocks and the request 1153 in total. The previous
+    # formula normalized both groups' memory by the first group's page size,
+    # reporting 3459/1057 = 3.27 here instead of 3 — and a different value
+    # again for the scheduler-config shape below.
+    uniform_full_spec = UniformTypeKVCacheSpecs(
+        block_size=full_attention_spec.block_size,
+        kv_cache_specs={f"layer_{i}": full_attention_spec for i in range(4)},
+    )
+    kv_cache_config_uniform_group = KVCacheConfig(
+        num_blocks=1153 * 3,
+        kv_cache_tensors=[],
+        kv_cache_groups=[
+            KVCacheGroupSpec([f"layer_{i}" for i in range(4)], uniform_full_spec),
+            KVCacheGroupSpec(["layer_4", "layer_5"], sliding_window_spec),
+        ],
+    )
+    assert (
+        get_max_concurrency_for_kv_cache_config(
+            vllm_config, kv_cache_config_uniform_group
+        )
+        == 3
+    )
+
+    # Scheduler-config shape: generate_scheduler_kv_cache_config replaces the
+    # uniform-type group's spec with a representative per-layer spec.
+    # Capacity must not change between the two shapes (the engine computes
+    # on the scheduler config, the worker loop on the worker config).
+    kv_cache_config_scheduler_shape = generate_scheduler_kv_cache_config(
+        [copy.deepcopy(kv_cache_config_uniform_group)]
+    )
+    assert get_max_concurrency_for_kv_cache_config(
+        vllm_config, kv_cache_config_scheduler_shape
+    ) == get_max_concurrency_for_kv_cache_config(
+        vllm_config, kv_cache_config_uniform_group
+    )
+
+
+def test_get_max_concurrency_packed_kv_cache_config():
+    from vllm.v1.core.kv_cache_utils import (
+        _get_kv_cache_config_packed,
+        _use_packed_kv_cache_config,
+    )
+
+    model_config = ModelConfig(
+        "Qwen/Qwen1.5-7B",
+        runner="generate",
+        dtype="float16",
+        max_model_len=16384,
+    )
+    scheduler_config = SchedulerConfig(
+        max_num_batched_tokens=1024,
+        enable_chunked_prefill=True,
+        max_model_len=model_config.max_model_len,
+        is_encoder_decoder=model_config.is_encoder_decoder,
+        async_scheduling=False,
+    )
+    vllm_config = VllmConfig(
+        model_config=model_config,
+        scheduler_config=scheduler_config,
+    )
+
+    # All-UniformTypeKVCacheSpecs groups select the packed layout.
+    mla_specs = {f"layer_{i}": new_mla_spec() for i in range(4)}
+    swa_specs = {
+        f"layer_{i}": SlidingWindowMLASpec(
+            block_size=16,
+            num_kv_heads=1,
+            head_size=576,
+            dtype=torch.float32,
+            sliding_window=128,
+        )
+        for i in range(4, 6)
+    }
+    kv_cache_groups = [
+        KVCacheGroupSpec(
+            list(mla_specs),
+            UniformTypeKVCacheSpecs(block_size=16, kv_cache_specs=mla_specs),
+        ),
+        KVCacheGroupSpec(
+            list(swa_specs),
+            UniformTypeKVCacheSpecs(block_size=16, kv_cache_specs=swa_specs),
+        ),
+    ]
+    assert _use_packed_kv_cache_config(vllm_config, kv_cache_groups)
+    num_blocks, kv_cache_tensors = _get_kv_cache_config_packed(
+        vllm_config, kv_cache_groups, 2 * GiB_bytes
+    )
+    assert num_blocks > 0
+    kv_cache_config_packed = KVCacheConfig(
+        num_blocks=num_blocks,
+        kv_cache_tensors=kv_cache_tensors,
+        kv_cache_groups=kv_cache_groups,
+    )
+    # Per-request blocks: the MLA group needs cdiv(16384, 16) = 1024 pages;
+    # the SWA group cdiv(min(128 - 1 + 1024, 16384), 16) + 1 = 73. The
+    # previous formula normalized by the first group's page size and gave
+    # 1061 blocks per request instead of 1097.
+    assert get_max_concurrency_for_kv_cache_config(
+        vllm_config, kv_cache_config_packed
+    ) == num_blocks / (1024 + 73)
+

 def test_allocate_with_lookahead():
    """Verify that lookahead tokens correctly affect block allocation"""
@@ -49,18 +49,6 @@ def test_prefix_caching_from_cli():
        args = parser.parse_args(["--prefix-caching-hash-algo", "invalid"])


-def test_extensible_kv_cache_from_cli():
-    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
-
-    args = parser.parse_args([])
-    engine_args = EngineArgs.from_cli_args(args=args)
-    assert not engine_args.enable_extensible_kv_cache
-
-    args = parser.parse_args(["--enable-extensible-kv-cache"])
-    engine_args = EngineArgs.from_cli_args(args=args)
-    assert engine_args.enable_extensible_kv_cache
-
-
@pytest.mark.skipif(_xxhash is None, reason="xxhash not installed")
 def test_prefix_caching_xxhash_from_cli():
    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Bugen ZhaoandOpenAI Codex	e1a763558c	[CI] Discover Rust coverage artifacts from build metadata Co-authored-by: OpenAI Codex <codex@openai.com>	2026-07-23 06:44:16 +00:00
Bugen ZhaoandOpenAI Codex	84aeec9f22	[CI] Simplify Rust coverage reporting Co-authored-by: OpenAI Codex <codex@openai.com>	2026-07-22 08:18:14 +00:00
Bugen ZhaoandOpenAI Codex	82a770ddbd	[CI] Simplify Rust coverage aggregation Co-authored-by: OpenAI Codex <codex@openai.com>	2026-07-22 02:58:45 +00:00
Bugen Zhao	a09a9bace1	[CI] Disable redundant Codecov file fixes	2026-07-21 13:57:38 +00:00
Bugen Zhao	6c20d467a2	[CI] Run Codecov from repository root	2026-07-21 13:34:22 +00:00
Bugen ZhaoandOpenAI Codex	cb59d0a351	[CI] Collect Rust coverage in Buildkite Co-authored-by: OpenAI Codex <codex@openai.com>	2026-07-21 13:01:56 +00:00
Bugen ZhaoandOpenAI Codex	0ab1bded36	[CI] Instrument Rust artifacts for coverage Co-authored-by: OpenAI Codex <codex@openai.com>	2026-07-21 12:22:25 +00:00
Umut PolatandGitHub	040cbf95cc	[Misc] Use VLLMValidationError in chat completion tool and batch validators (#49214 ) Signed-off-by: Umut Polat <52835619+umut-polat@users.noreply.github.com>	2026-07-21 11:38:20 +00:00
Sya GitHubLi, Jiang <jiang1.li@intel.com>	5b3762a7f0	[Bugfix][CPU] Fix Clang OpenMP build on macOS (#49021 ) Signed-off-by: markyangcc <mmdou3@163.com> Co-authored-by: Li, Jiang <jiang1.li@intel.com>	2026-07-21 09:58:52 +00:00
bastefaniakandGitHub	4d30c510ce	[bugfix] Fix Cosmos3 Edge checkpoint weights filtering, video loading, prompt expansion (#49190 ) Signed-off-by: Bartosz Stefaniak <bstefaniak@nvidia.com>	2026-07-21 17:18:36 +08:00
Lucas Wilkinson GitHub Claude OpenAI Codex	6700813f86	[3/N][KV-Cache Layout Refactor] Standardize Mamba cache; drop `get_transfer_cache_regions` (#44456 ) Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: OpenAI Codex <codex@openai.com>	2026-07-21 09:16:15 +00:00
Bugen ZhaoandGitHub	eb44b3aaa4	[Rust][Benchmark] Use async HTTP clients (#49295 ) Signed-off-by: Bugen Zhao <i@bugenzhao.com>	2026-07-21 16:53:57 +08:00
Nicolò LucchesiandGitHub	7a98c7a392	[Misc] Remove old now unsupported `max_num_partial_prefills` and `max_long_partial_prefills` (#49244 ) Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>	2026-07-21 08:52:52 +00:00
Lena OnyshchenkoandGitHub	0d9e60619b	[Misc][Docs] Fix XPU compute-runtime driver link version mismatch (#49299 ) Signed-off-by: oonyshch <xonyshch@gmail.com>	2026-07-21 08:45:41 +00:00
vllm-agent GitHub vllm-agent	1134545b6f	Revert "[Sampler] Stop upcasting logits to fp32 in apply_sampling_params" (#48641 ) (#49033 ) Co-authored-by: vllm-agent <vllm-agent@users.noreply.github.com>	2026-07-21 09:36:45 +01:00
Miłosz Grunwald GitHubmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>Harry Mellor	3e0c887511	[Bugfix] Fix Ovis2_5 special tokens for transformers v5 (#47298 ) Signed-off-by: mgrunwal <milosz.grunwald@intel.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>	2026-07-21 08:09:20 +00:00
Stefan KaestleandGitHub	adfbbc1005	Propagate Flash Attention cache configuration to Ray workers (#49177 ) Signed-off-by: Stefan Kaestle <skaestle@nvidia.com>	2026-07-21 07:47:53 +00:00
Roy WangandGitHub	adc98f04d0	[Misc] Add @esmeetu to codeowners for rust/src/bench (#49298 ) Signed-off-by: esmeetu <jasonailu87@gmail.com>	2026-07-21 07:44:24 +00:00
Alex Yuan GitHub Claude Opus 4.8	8def3cdde2	[Bugfix] Propagate quant_config to LFM2 ShortConv projections (#48917 ) Signed-off-by: Alex Yuan <alex.yuan@liquid.ai> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-21 06:51:55 +00:00
Guan-Ming Chiu GitHub Cyrus Leung	616c9bd0f4	[Frontend] Support additional sampling parameters for translation API (#45839 ) Signed-off-by: Guan-Ming (Wesley) Chiu <105915352+guan404ming@users.noreply.github.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>	2026-07-21 05:53:56 +00:00
Bugen ZhaoandGitHub	8688a06d67	[Rust][Benchmark] Use `tracing` for logs (#48937 ) Signed-off-by: Bugen Zhao <i@bugenzhao.com>	2026-07-21 05:18:13 +00:00
Salt Sato GitHubmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	f25953cc59	[Bugfix][Rust Frontend] Handle zero-column logprobs payloads without panicking (#49113 ) Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Signed-off-by: Feathbow <feathbow@gmail.com>	2026-07-21 04:30:29 +00:00
aoshen02 GitHub aoshen02 Codex	d9aa35161d	Update BGE-M3 token expectations for leading spaces (#49269 ) Signed-off-by: aoshen02 <aoshen02@users.noreply.github.com> Co-authored-by: aoshen02 <aoshen02@users.noreply.github.com> Co-authored-by: Codex <noreply@openai.com>	2026-07-21 03:49:17 +00:00
Andreas Karatzas GitHub OpenAI Codex	6bcda970fd	[CI][NIXL] Isolate concurrent engine internal ports (#49129 ) Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com> Co-authored-by: OpenAI Codex <noreply@openai.com>	2026-07-20 22:28:11 -05:00
Isotr0pyandGitHub	ea0e9c8f2e	[MRV2] Add encoder cache profiling implementation (#47985 ) Signed-off-by: Isotr0py <Isotr0py@outlook.com>	2026-07-20 20:18:14 -07:00
ChaunceyandGitHub	94ed0bf4e0	[Bugfix][KV Offloading] Handle queued request aborts without allocated KV blocks (#49146 ) Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>	2026-07-21 11:16:26 +08:00
Connor Carpenter GitHubmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	1940c8441e	[Rust Frontend][gRPC] Add engine-aware health reporting (#48992 ) Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Signed-off-by: Connor Carpenter <connorc@nvidia.com>	2026-07-21 10:54:39 +08:00
Simon MoandGitHub	72d16aee15	[CI] Exercise FA3 FP8 attention on SM90 (#49231 ) Signed-off-by: Simon Mo <simon@inferact.ai>	2026-07-21 10:26:55 +08:00
Kunshang JiandGitHub	e78a0c8e59	[XPU][Doc] Update XPU docker image documents (#49148 ) Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>	2026-07-21 10:11:27 +08:00
Chris LeonardandGitHub	97a98006b0	Update qutlass cmake for stable abi (#47879 ) Signed-off-by: Chris Leonard <chleonar@redhat.com>	2026-07-20 18:31:10 -07:00
Alejandro Paredes La Torre GitHub Shengqi Chen	0a684ab0c0	[Bugfix] Fix WSL circular import from pin_memory warning_once (#48444 ) Signed-off-by: AlejandroParedesLT <alejandroparedeslatorre@gmail.com> Co-authored-by: Shengqi Chen <harry-chen@outlook.com>	2026-07-20 18:30:55 -07:00
flutist-alibaba GitHub Copilot Autofix powered by AI Shengqi Chen	0d9210a502	Fixes non-coalesced HBM access in marlin_int4_fp8_preprocess_kernel_awq (#47268 ) Signed-off-by: xjx <493337577@qq.com> Signed-off-by: flutist-alibaba <30485581+flutist@users.noreply.github.com> Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> Co-authored-by: Shengqi Chen <harry-chen@outlook.com>	2026-07-20 18:30:38 -07:00
Lena Onyshchenko GitHub Cursor	1d874867ea	[Misc][Docs] Fix broken protocol link in speech_to_text doc (#47212 ) Signed-off-by: oonyshch <xonyshch@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com>	2026-07-21 01:05:05 +00:00
ormandj GitHub Luke Alonso Martin Vit Claude Yifan Qiao	2e2e626b40	[Bugfix] Count per-group blocks in get_max_concurrency_for_kv_cache_config (#48317 ) Signed-off-by: David Orman <ormandj@corenode.com> Co-authored-by: Luke Alonso <lalonso@gmail.com> Co-authored-by: Martin Vit <martin@voipmonitor.org> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>	2026-07-21 00:29:10 +00:00
Nick HillandGitHub	af91f4b3e4	[Cleanup] Remove unused StructuredOutputRequest.status field (#49235 ) Signed-off-by: Nick Hill <nickhill123@gmail.com>	2026-07-21 00:06:16 +00:00
sungsoo ha GitHub Claude Opus 4.8 Lucas Wilkinson Matthew Bonanni	2396a61108	[Attention][MLA][DCP] Query replication for MLA decode (DeepSeek-V2/R1 + Kimi-K2.5) (#45964 ) Signed-off-by: Sungsoo Ha <sungsooh@nvidia.com> Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com> Signed-off-by: Matthew Bonanni <mbonanni@redhat.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com> Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>	2026-07-20 23:51:27 +00:00
Xuanyu Zhang GitHub Claude aoshen02 OpenAI Codex	97a668152b	[RL Infra][FlashInfer] Enable router replay output from FlashInfer monolithic MoE kernel (#44214 ) Signed-off-by: Xuanyu Zhang <xuanyu.zhang@mistral.ai> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: aoshen02 <aoshen@inferact.ai> Co-authored-by: OpenAI Codex <noreply@openai.com>	2026-07-20 16:45:10 -07:00
Matthew Bonanni GitHub haosdent Claude Opus 4.7 Justin Wood	58b2012aa2	[copy of #45208 ] CuMem slept-L1 fragmentation accounting (#49208 ) Signed-off-by: haosdent <haosdent@gmail.com> Signed-off-by: Justin Wood <justin.m.wood@me.com> Signed-off-by: Matthew Bonanni <mbonanni@redhat.com> Co-authored-by: haosdent <haosdent@gmail.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: Justin Wood <jwood@me.com>	2026-07-20 23:08:11 +00:00
Ning XieandGitHub	b7c20d0cfa	[chore] adjust logo be more friendly to white background terminal (#48938 ) Signed-off-by: Andy Xie <andy.xning@gmail.com>	2026-07-20 15:15:28 -07:00
TJianandGitHub	a2b1f9fc3b	[ROCm] [Release] [Bugfix] Fix the per commit wheel release pipeline. (#49245 ) Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>	2026-07-20 22:12:38 +00:00
Fynn Schmitt-Ulms GitHub Michael Goin	642076d26c	Support loading sample_from_anchor flag from speculators config (#48639 ) Signed-off-by: Fynn Schmitt-Ulms <fschmitt@redhat.com> Co-authored-by: Michael Goin <mgoin64@gmail.com>	2026-07-20 14:52:43 -07:00
Charlie FuandGitHub	5feb3950e5	[ROCm][CI] fix test_rocm_quick_reduce.py (#49234 ) Signed-off-by: charlifu <charlifu@amd.com>	2026-07-20 16:39:33 -05:00
Oxana Korzh GitHub Cursor okorzh-amd Matthew Bonanni	4ec199b66a	[Bugfix][Spec-Decode] Populate draft seq_lens_cpu_upper_bound for spec-decode attention metadata (#44492 ) Signed-off-by: Oxana Korzh <okorzh@amd.com> Signed-off-by: okorzh-amd <okorzh-amd@users.noreply.github.com> Signed-off-by: Matthew Bonanni <mbonanni@redhat.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: okorzh-amd <okorzh-amd@users.noreply.github.com> Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>	2026-07-20 20:50:58 +00:00
Roberto L. Castro GitHubmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	7ca017778f	[Feat][Perf] Add new warmup infrastructure for JITs (#47451 ) Signed-off-by: LopezCastroRoberto <rocastro@redhat.com> Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	2026-07-20 13:21:55 -07:00
coltonottley GitHub Colton Ottley Or Ozeri	fbfe58133d	[Bugfix][KV Offload] Preserve reachable tails for hybrid SWA groups (#48911 ) Signed-off-by: Colton Ottley <colton@ottleyengineering.com> Co-authored-by: Colton Ottley <colton@ottleyengineering.com> Co-authored-by: Or Ozeri <oro@il.ibm.com>	2026-07-20 22:12:36 +03:00
wkutak GitHub Isotr0py Roger Wang	9dd62d80ab	Cosmos3 FP8 ModelOpt/Diffusers remapping (#48952 ) Signed-off-by: Wojciech Kutak <wkutak@nvidia.com> Signed-off-by: wkutak <wkutak@nvidia.com> Signed-off-by: Isotr0py <Isotr0py@outlook.com> Co-authored-by: Isotr0py <Isotr0py@outlook.com> Co-authored-by: Roger Wang <hey@rogerw.io>	2026-07-20 11:31:48 -07:00
tc-mb GitHub wjinxu	f878367898	[Revert][Bugfix] Restore MiniCPM-V 4.6 ViT QKV weight loader (#49193 ) Signed-off-by: wjinxu <1299461899@qq.com> Co-authored-by: wjinxu <1299461899@qq.com>	2026-07-20 18:17:46 +00:00
Matthew Bonanni GitHubmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	bd091079cb	[Attention] FlashAttention 4 SM100 FP8 kv cache support (#42569 ) Signed-off-by: Matthew Bonanni <mbonanni@redhat.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	2026-07-20 10:53:27 -07:00
Liangqiusong GitHub Kunshang Ji	b23bd73f54	[XPU]add sycl path for Mhc (#47245 ) Signed-off-by: root <xiaolong.guo@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>	2026-07-20 15:32:54 +00:00
Bugen ZhaoandGitHub	e2d7adeb64	[Rust Frontend] Bump `xgrammar-structural-tag` and enable local extension (#49161 ) Signed-off-by: Bugen Zhao <i@bugenzhao.com>	2026-07-20 16:22:24 +01:00
Isotr0pyandGitHub	15cb8e140d	[Multimodal] Allow keeping original image mode for ImageIO (#49159 ) Signed-off-by: Isotr0py <Isotr0py@outlook.com>	2026-07-20 13:42:45 +00:00
Chang Guo GitHub OpenAI Codex	f007cceb42	[KV Offload] Support self-describing KV events with TieringOffloadingSpec (#48679 ) Signed-off-by: Change72 <changg@nvidia.com> Co-authored-by: OpenAI Codex <codex@openai.com>	2026-07-20 16:41:58 +03:00
wangqian GitHub OpenAI Codex	0a5069e4e3	[Bugfix][Gemma4] Fix ModelOpt mixed-precision MoE config mapping (#48563 ) Signed-off-by: wangqian <601731555@qq.com> Co-authored-by: OpenAI Codex <codex@openai.com>	2026-07-20 06:39:28 -07:00
EdalatiAli GitHub Cursor Nicolò Lucchesi	8ce53a616e	[Bugfix] Zero new KV blocks for quantized + sliding-window hybrid caches (#47574 ) Signed-off-by: EdalatiAli <aliedalati@cohere.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>	2026-07-20 13:18:17 +00:00