Fix AMD standardized KV cache regressions

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

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>

.buildkite/hardware_tests/cpu.yaml

@@ -18,6 +18,8 @@ steps:
   - tests/kernels/quantization/test_cpu_fp8_scaled_mm.py
   - tests/kernels/mamba/cpu/test_cpu_gdn_ops.py
   - tests/kernels/mamba/test_cpu_short_conv.py
+  - tests/kernels/mamba/test_causal_conv1d.py
+  - tests/kernels/mamba/test_mamba_ssm.py
   commands:
     - |
       bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
@@ -28,7 +30,9 @@ steps:
       pytest -x -v -s tests/kernels/test_onednn.py
       pytest -x -v -s tests/kernels/test_awq_int4_to_int8.py
       pytest -x -v -s tests/kernels/quantization/test_cpu_fp8_scaled_mm.py
-      pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py"
+      pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py
+      pytest -x -v -s tests/kernels/mamba/test_causal_conv1d.py
+      pytest -x -v -s tests/kernels/mamba/test_mamba_ssm.py"
 
 # Note: SDE can't be downloaded from CI host because of AWS WAF
 # - label: CPU-Compatibility Tests

.buildkite/release-pipeline.yaml

@@ -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

.buildkite/scripts/hardware_ci/run-cpu-test-arm.sh

@@ -40,7 +40,9 @@ function cpu_tests() {
     pytest -x -v -s tests/kernels/moe/test_cpu_fused_moe.py
     pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py
     pytest -x -v -s tests/kernels/moe/test_cpu_int4_moe.py
-    pytest -x -v -s tests/kernels/mamba/test_cpu_short_conv.py"
+    pytest -x -v -s tests/kernels/mamba/test_cpu_short_conv.py
+    pytest -x -v -s tests/kernels/mamba/test_causal_conv1d.py
+    pytest -x -v -s tests/kernels/mamba/test_mamba_ssm.py"
 
   # skip tests requiring model downloads if HF_TOKEN is not set
   # due to rate-limits
@@ -97,3 +99,4 @@ function cpu_tests() {
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
 timeout 2h bash -c cpu_tests
+

.buildkite/test_areas/model_executor.yaml

@@ -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

.github/CODEOWNERS

@@ -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

README.md

@@ -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:
 

benchmarks/attention_benchmarks/runner.py

@@ -12,6 +12,7 @@ import logging
 import statistics
 import types
 from contextlib import contextmanager
+from math import prod
 
 import torch
 from batch_spec import parse_batch_spec, reorder_for_flashinfer
@@ -37,10 +38,13 @@ from vllm.config import (
 )
 from vllm.v1.attention.backends.utils import (
     CommonAttentionMetadata,
-    get_kv_cache_layout,
-    set_kv_cache_layout,
+    resolve_kv_cache_layout,
+)
+from vllm.v1.kv_cache_interface import (
+    FullAttentionSpec,
+    compute_layer_kv_cache_shape_bytes,
+    reshape_kv_cache,
 )
-from vllm.v1.kv_cache_interface import FullAttentionSpec
 
 # ============================================================================
 # Backend Configuration
@@ -337,52 +341,23 @@ def _create_input_tensors(
 def _create_kv_cache(
     config: BenchmarkConfig,
     max_num_blocks: int,
-    backend_class,
     device: torch.device,
     dtype: torch.dtype,
 ) -> list:
-    """Create KV cache tensors for all layers using the backend's methods.
-
-    Uses the backend's get_kv_cache_shape() and get_kv_cache_stride_order()
-    to create the cache with the correct shape and memory layout.
-    """
-    # Get the logical shape from the backend
-    cache_shape = backend_class.get_kv_cache_shape(
-        num_blocks=max_num_blocks,
+    """Create KV cache tensors for all layers using the standard allocator."""
+    spec = FullAttentionSpec(
         block_size=config.block_size,
         num_kv_heads=config.num_kv_heads,
         head_size=config.head_dim,
+        dtype=dtype,
     )
-
-    # Get the stride order for custom memory layout
-    try:
-        stride_order = backend_class.get_kv_cache_stride_order()
-        assert len(stride_order) == len(cache_shape)
-    except (AttributeError, NotImplementedError):
-        stride_order = tuple(range(len(cache_shape)))
-
-    # Permute shape to physical layout order
-    physical_shape = tuple(cache_shape[i] for i in stride_order)
-
-    # Compute inverse permutation to get back to logical view
-    inv_order = [stride_order.index(i) for i in range(len(stride_order))]
-
-    # Use fp8 dtype for cache when requested.
-    cache_dtype = dtype
-    if config.kv_cache_dtype == "fp8":
-        from vllm.platforms import current_platform
-
-        cache_dtype = current_platform.fp8_dtype()
-
-    cache_list = []
-    for _ in range(config.num_layers):
-        # Allocate in physical layout order (contiguous in memory)
-        cache = torch.zeros(*physical_shape, device=device, dtype=cache_dtype)
-        # Permute to logical view
-        cache = cache.permute(*inv_order)
-        cache_list.append(cache)
-
-    return cache_list
+    layout = resolve_kv_cache_layout()
+    total_bytes = (
+        prod(compute_layer_kv_cache_shape_bytes(spec, max_num_blocks))
+        * config.num_layers
+    )
+    buf = torch.zeros(total_bytes, device=device, dtype=torch.int8)
+    return reshape_kv_cache(buf, spec, max_num_blocks, config.num_layers, layout)
 
 
 # ============================================================================
@@ -500,13 +475,6 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 backend_cfg, config, device, dtype
             )
 
-            # Set KV cache layout if the backend requires a specific one
-            # (e.g., FlashInfer requires HND on SM100/Blackwell for TRTLLM attention)
-            required_layout = backend_class.get_required_kv_cache_layout()
-            if required_layout is not None:
-                set_kv_cache_layout(required_layout)
-                get_kv_cache_layout.cache_clear()
-
             common_metadata = _build_common_attn_metadata(
                 q_lens, kv_lens, config.block_size, device
             )
@@ -541,9 +509,7 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 config, total_q, device, dtype, quantize_query=quantize_query
             )
 
-            cache_list = _create_kv_cache(
-                config, max_num_blocks, backend_class, device, dtype
-            )
+            cache_list = _create_kv_cache(config, max_num_blocks, device, dtype)
 
             timing_stats, mem_stats = _run_single_benchmark(
                 config,

cmake/cpu_extension.cmake

@@ -430,6 +430,7 @@ set(VLLM_EXT_SRC
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
+    "csrc/cpu/mamba_cpu.cpp"
     "csrc/moe/dynamic_4bit_int_moe_cpu.cpp"
     "csrc/cpu/cpu_attn.cpp"
     "csrc/cpu/torch_bindings.cpp")
@@ -489,6 +490,7 @@ if (ENABLE_X86_ISA)
         "csrc/cpu/spec_decode_utils.cpp"
         "csrc/cpu/cpu_attn.cpp"
         "csrc/cpu/dnnl_kernels.cpp"
+        "csrc/cpu/mamba_cpu.cpp"
         "csrc/cpu/torch_bindings.cpp"
         # TODO: Remove these files
         "csrc/cpu/activation.cpp"
@@ -502,6 +504,7 @@ if (ENABLE_X86_ISA)
         "csrc/cpu/utils.cpp"
         "csrc/cpu/spec_decode_utils.cpp"
         "csrc/cpu/cpu_attn.cpp"
+        "csrc/cpu/mamba_cpu.cpp"
         "csrc/cpu/dnnl_kernels.cpp"
         "csrc/cpu/torch_bindings.cpp"
         # TODO: Remove these files

cmake/external_projects/qutlass.cmake

@@ -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>

cmake/external_projects/vllm_flash_attn.cmake

@@ -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

csrc/cpu/cpu_types_vsx.hpp

@@ -336,13 +336,14 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
     reg.val[1] = fp16_to_fp32_bits(raw_lo);
   }
   float reduce_sum() const {
-    AliasReg ar;
-    ar.reg = reg;
-    float result = 0;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&result, &ar](int i) { result += ar.values[i]; });
-
-    return result;
+    // VSX horizontal reduction: 3 vector ops instead of 8 scalar adds.
+    // Step 1: pairwise sum of the two 4-wide halves
+    __vector float s = vec_add(reg.val[0], reg.val[1]);
+    // Step 2: rotate by 8 bytes (2 floats) and add
+    s = vec_add(s, vec_sld(s, s, 8));
+    // Step 3: rotate by 4 bytes (1 float) and add  => all lanes hold total
+    s = vec_add(s, vec_sld(s, s, 4));
+    return vec_extract(s, 0);
   }
   FP32Vec8 exp() const {
     f32x4x2_t out;

csrc/cpu/mamba_cpu.cpp

@@ -0,0 +1,285 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+//
+// CPU at::Tensor wrappers for Mamba decode-step kernels defined in
+// mamba_kernels.hpp.
+
+#include "cpu/mamba_kernels.hpp"
+
+#include <ATen/ATen.h>
+#include <torch/library.h>
+#include <c10/util/Optional.h>
+
+#include "cpu_types.hpp"
+
+// ---------------------------------------------------------------------------
+// causal_conv1d_update
+// ---------------------------------------------------------------------------
+at::Tensor causal_conv1d_update_cpu_impl(
+    at::Tensor& x, at::Tensor& conv_state, const at::Tensor& weight,
+    const c10::optional<at::Tensor>& bias,
+    const c10::optional<std::string>& activation,
+    const c10::optional<at::Tensor>& conv_state_indices,
+    const c10::optional<at::Tensor>& query_start_loc, int64_t pad_slot_id) {
+  bool do_silu = false;
+  if (activation.has_value()) {
+    const std::string& act = activation.value();
+    do_silu = (act == "silu" || act == "swish");
+  }
+
+  at::ScalarType dtype = x.scalar_type();
+
+  // Input x: contiguous in native dtype.
+  at::Tensor x_c = x.is_contiguous() ? x : x.contiguous();
+
+  // conv_state: NEVER copy the full paged tensor just for layout reasons.
+  // If the dtype matches we work directly on conv_state (contiguous or not)
+  // by extracting strides and passing them to the kernel.
+  // Only a dtype-conversion copy is made when types differ (rare for BF16).
+  bool state_type_ok = (conv_state.scalar_type() == dtype);
+  at::Tensor state_c = state_type_ok ? conv_state : conv_state.to(dtype);
+  // state_c and conv_state may be non-contiguous — that is intentional.
+
+  // Weight: coerce to same dtype if needed (should match in practice)
+  at::Tensor w_c =
+      (weight.scalar_type() != dtype)
+          ? weight.to(dtype).contiguous()
+          : (weight.is_contiguous() ? weight : weight.contiguous());
+
+  // Bias stays float32 (small scalar, used only for fp32 accumulation)
+  at::Tensor bias_f32;
+  if (bias.has_value() && bias.value().defined())
+    bias_f32 = bias.value().to(at::kFloat).contiguous();
+
+  int64_t batch = x_c.size(0);
+  int64_t dim = x_c.size(1);
+  int64_t seqlen = (x_c.dim() == 3) ? x_c.size(2) : 1;
+  int64_t width = w_c.size(1);
+  int64_t state_len = state_c.size(2);
+
+  // Extract strides — works for contiguous AND non-contiguous (transposed)
+  // state. stride(0): between cache slots (e.g. num_slots × dim × width-1 in
+  // contiguous) stride(1): between conv channels (dim stride) stride(2):
+  // between state elements (=1 when contiguous, =dim when transposed)
+  int64_t stride_s_slot = state_c.stride(0);
+  int64_t stride_s_dim = state_c.stride(1);
+  int64_t stride_s_state = state_c.stride(2);
+
+  at::Tensor out = x_c.clone();  // native dtype, no float32 alloc
+
+  const int32_t* cache_idx_ptr = nullptr;
+  at::Tensor cache_idx_int;
+  if (conv_state_indices.has_value()) {
+    cache_idx_int = conv_state_indices.value().to(at::kInt).contiguous();
+    cache_idx_ptr = cache_idx_int.data_ptr<int32_t>();
+  }
+
+  VLLM_DISPATCH_FLOATING_TYPES(dtype, "causal_conv1d_update", [&] {
+    mamba_cpu::causal_conv1d_update_kernel<scalar_t>(
+        x_c.data_ptr<scalar_t>(), state_c.data_ptr<scalar_t>(), stride_s_slot,
+        stride_s_dim, stride_s_state, w_c.data_ptr<scalar_t>(),
+        bias_f32.defined() ? bias_f32.data_ptr<float>() : nullptr,
+        out.data_ptr<scalar_t>(), cache_idx_ptr,
+        static_cast<int32_t>(pad_slot_id), batch, dim, seqlen, width, state_len,
+        do_silu);
+  });
+
+  // Write back only when a type-conversion copy was made.
+  // Layout-only non-contiguity is handled via strides above — no copy needed.
+  if (!state_type_ok) conv_state.copy_(state_c);
+
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// selective_state_update
+// ---------------------------------------------------------------------------
+void selective_state_update_cpu_impl(
+    at::Tensor& state,    // (nstates, nheads, dim, dstate)
+    const at::Tensor& x,  // (N, nheads, dim)
+    const at::Tensor& dt, const at::Tensor& A, const at::Tensor& B,
+    const at::Tensor& C, const c10::optional<at::Tensor>& D,
+    const c10::optional<at::Tensor>& z,
+    const c10::optional<at::Tensor>& dt_bias, bool dt_softplus,
+    const c10::optional<at::Tensor>& state_batch_indices,
+    const c10::optional<at::Tensor>& dst_state_batch_indices,
+    int64_t null_block_id, at::Tensor& out,
+    const c10::optional<at::Tensor>& num_accepted_tokens,
+    const c10::optional<at::Tensor>& cu_seqlens) {
+  at::ScalarType state_type = state.scalar_type();
+  at::ScalarType input_type = x.scalar_type();
+
+  // x, B, C must be contiguous and match input_type
+  auto ensure_input = [input_type](const at::Tensor& t) -> at::Tensor {
+    at::Tensor r = (t.scalar_type() != input_type) ? t.to(input_type) : t;
+    return r.is_contiguous() ? r : r.contiguous();
+  };
+  at::Tensor x_in = ensure_input(x);
+  at::Tensor B_in = ensure_input(B);
+  at::Tensor C_in = ensure_input(C);
+  at::Tensor z_in;
+  if (z.has_value() && z.value().defined()) z_in = ensure_input(z.value());
+
+  // A, D, dt_bias are float32 model parameters that arrive here as expanded
+  // tensors, e.g. A is (nheads, head_dim, dstate) with strides (1, 0, 0).
+  // We need just the scalar value per head as a (nheads,) 1-D array so that
+  // A_ptr[h] in the kernel correctly reads head h's value.
+  //
+  // Strategy: peel trailing expanded (stride=0) dims via .select(), which is
+  // a zero-copy view.  For A: (nheads, head_dim, dstate) strides (1,0,0)
+  //   → .select(2,0) → (nheads, head_dim) strides (1,0)
+  //   → .select(1,0) → (nheads,)          stride  (1,)   ← contiguous, free.
+  // No allocation, no type conversion (A is already float32).
+  auto to_per_head_1d_f32 = [](const at::Tensor& t) -> at::Tensor {
+    at::Tensor r = t;
+    // Peel trailing dimensions that are broadcast (stride=0 or size=1)
+    while (r.dim() > 1) r = r.select(r.dim() - 1, 0);
+    if (r.scalar_type() != at::kFloat) r = r.to(at::kFloat);
+    return r.is_contiguous() ? r : r.contiguous();
+  };
+
+  at::Tensor A_f32 = to_per_head_1d_f32(A);  // (nheads,) float32
+  at::Tensor D_f32, dt_bias_f32;
+  if (D.has_value() && D.value().defined())
+    D_f32 = to_per_head_1d_f32(D.value());
+  if (dt_bias.has_value() && dt_bias.value().defined())
+    dt_bias_f32 = to_per_head_1d_f32(dt_bias.value());
+
+  // dt: reduce (N, nheads, head_dim) expanded tensor → (N, nheads) BEFORE
+  // the type conversion so we convert head_dim x fewer elements.
+  at::Tensor dt_f32;
+  {
+    // If dt was expanded to (N, nheads, head_dim) with stride-0 in dim 2,
+    // take a zero-copy view of index 0 along that dim first.
+    at::Tensor t2 = (dt.dim() == 3) ? dt.select(2, 0) : dt;  // (N, nheads)
+    at::Tensor t3 = (t2.scalar_type() != at::kFloat) ? t2.to(at::kFloat) : t2;
+    dt_f32 = t3.is_contiguous() ? t3 : t3.contiguous();
+  }
+
+  int64_t nheads = state.size(1);
+  int64_t dim = state.size(2);
+  int64_t dstate = state.size(3);
+  int64_t N = (cu_seqlens.has_value() && cu_seqlens.value().defined())
+                  ? cu_seqlens.value().size(0) - 1
+                  : x_in.size(0);
+  int64_t ngroups = B_in.size(1);
+
+  // Strides
+  int64_t stride_state_n = state.stride(0);
+  int64_t stride_state_h = state.stride(1);
+  int64_t stride_state_d = state.stride(2);
+  int64_t stride_x_n = x_in.stride(0);
+  int64_t stride_x_h = x_in.stride(1);
+  int64_t stride_dt_n = dt_f32.stride(0);  // dt is (N, nheads)
+  int64_t stride_BC_n = B_in.stride(0);
+  int64_t stride_BC_g = B_in.stride(1);
+  int64_t stride_out_n = out.stride(0);
+  int64_t stride_out_h = out.stride(1);
+
+  // Optional index pointers
+  auto get_int32_ptr =
+      [](const c10::optional<at::Tensor>& opt) -> const int32_t* {
+    return (opt.has_value() && opt.value().defined())
+               ? opt.value().data_ptr<int32_t>()
+               : nullptr;
+  };
+  const int32_t* sbi_ptr = get_int32_ptr(state_batch_indices);
+  const int32_t* dsbi_ptr = get_int32_ptr(dst_state_batch_indices);
+  const int32_t* nat_ptr = get_int32_ptr(num_accepted_tokens);
+  const int32_t* csl_ptr = get_int32_ptr(cu_seqlens);
+
+  // Dispatch on (state_t, input_t, out_t): write directly into `out`
+  // without any intermediate float32 buffer.
+  VLLM_DISPATCH_FLOATING_TYPES(state_type, "ssu_state", [&] {
+    using state_t = scalar_t;
+    VLLM_DISPATCH_FLOATING_TYPES(input_type, "ssu_input", [&] {
+      using input_t = scalar_t;
+      VLLM_DISPATCH_FLOATING_TYPES(out.scalar_type(), "ssu_out", [&] {
+        using out_t = scalar_t;
+        mamba_cpu::selective_state_update_kernel<state_t, input_t, out_t>(
+            state.data_ptr<state_t>(), stride_state_n, stride_state_h,
+            stride_state_d, x_in.data_ptr<input_t>(), stride_x_n, stride_x_h,
+            dt_f32.data_ptr<float>(), stride_dt_n, A_f32.data_ptr<float>(),
+            B_in.data_ptr<input_t>(), C_in.data_ptr<input_t>(), stride_BC_n,
+            stride_BC_g, D_f32.defined() ? D_f32.data_ptr<float>() : nullptr,
+            z_in.defined() ? z_in.data_ptr<input_t>() : nullptr,
+            dt_bias_f32.defined() ? dt_bias_f32.data_ptr<float>() : nullptr,
+            out.data_ptr<out_t>(), stride_out_n, stride_out_h, sbi_ptr,
+            dsbi_ptr, static_cast<int32_t>(null_block_id), nat_ptr, csl_ptr, N,
+            nheads, ngroups, dim, dstate, dt_softplus);
+      });
+    });
+  });
+}
+
+// ---------------------------------------------------------------------------
+// mamba_chunk_scan_fwd_cpu
+// ---------------------------------------------------------------------------
+void mamba_chunk_scan_fwd_cpu_impl(
+    at::Tensor& out,  // [seqlen, nheads, headdim] — pre-allocated by caller
+    at::Tensor&
+        final_states,     // [batch, nheads, headdim, dstate] float32 contiguous
+    const at::Tensor& x,  // [seqlen, nheads, headdim]
+    const at::Tensor&
+        dt,  // [seqlen, nheads] float32 (preprocessed: bias+softplus+clamp)
+    const at::Tensor& A,                 // [nheads] float32
+    const at::Tensor& B,                 // [seqlen, ngroups, dstate]
+    const at::Tensor& C,                 // [seqlen, ngroups, dstate]
+    const c10::optional<at::Tensor>& D,  // [nheads] float32 (optional)
+    const c10::optional<at::Tensor>& z,  // [seqlen, nheads, headdim] (optional)
+    const at::Tensor& cu_seqlens         // [batch+1] int32
+) {
+  const at::ScalarType input_type = x.scalar_type();
+
+  auto ensure_contig = [input_type](const at::Tensor& t) -> at::Tensor {
+    at::Tensor r = (t.scalar_type() != input_type) ? t.to(input_type) : t;
+    return r.is_contiguous() ? r : r.contiguous();
+  };
+  at::Tensor x_in = ensure_contig(x);
+  at::Tensor B_in = ensure_contig(B);
+  at::Tensor C_in = ensure_contig(C);
+  at::Tensor z_in;
+  if (z.has_value() && z.value().defined()) z_in = ensure_contig(z.value());
+
+  // A and D are float32 model parameters, potentially broadcast-expanded.
+  // Strip trailing broadcast dims to get a contiguous (nheads,) array.
+  auto to_per_head_f32 = [](const at::Tensor& t) -> at::Tensor {
+    at::Tensor r = t;
+    while (r.dim() > 1) r = r.select(r.dim() - 1, 0);
+    if (r.scalar_type() != at::kFloat) r = r.to(at::kFloat);
+    return r.is_contiguous() ? r : r.contiguous();
+  };
+  at::Tensor A_f32 = to_per_head_f32(A);
+  at::Tensor D_f32;
+  if (D.has_value() && D.value().defined()) D_f32 = to_per_head_f32(D.value());
+
+  // dt: [seqlen, nheads] float32 — caller has applied bias+softplus+clamp in
+  // Python.
+  at::Tensor dt_c = dt.is_contiguous() ? dt : dt.contiguous();
+  if (dt_c.scalar_type() != at::kFloat) dt_c = dt_c.to(at::kFloat);
+
+  at::Tensor cu_int = cu_seqlens.to(at::kInt).contiguous();
+
+  const int64_t batch = final_states.size(0);
+  const int64_t nheads = final_states.size(1);
+  const int64_t headdim = final_states.size(2);
+  const int64_t dstate = final_states.size(3);
+  const int64_t ngroups = B_in.size(1);
+
+  TORCH_CHECK(final_states.is_contiguous(),
+              "mamba_chunk_scan_fwd_cpu: final_states must be contiguous");
+  TORCH_CHECK(out.is_contiguous(),
+              "mamba_chunk_scan_fwd_cpu: out must be contiguous (writes via "
+              "raw data_ptr)");
+
+  VLLM_DISPATCH_FLOATING_TYPES(input_type, "mamba_chunk_scan_fwd_cpu", [&] {
+    mamba_cpu::mamba_chunk_scan_fwd_kernel<scalar_t>(
+        final_states.data_ptr<float>(), x_in.data_ptr<scalar_t>(),
+        dt_c.data_ptr<float>(), A_f32.data_ptr<float>(),
+        B_in.data_ptr<scalar_t>(), C_in.data_ptr<scalar_t>(),
+        D_f32.defined() ? D_f32.data_ptr<float>() : nullptr,
+        z_in.defined() ? z_in.data_ptr<scalar_t>() : nullptr,
+        out.data_ptr<scalar_t>(), cu_int.data_ptr<int32_t>(), batch, nheads,
+        ngroups, headdim, dstate);
+  });
+}

csrc/cpu/mamba_kernels.hpp

@@ -0,0 +1,382 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+//
+// Fused CPU vector kernels for Mamba decode-step hotspots:
+//   - causal_conv1d_update  (depthwise 1-D conv state roll + compute)
+//   - selective_state_update (SSM recurrence, single-step)
+
+#pragma once
+
+#include "cpu_types.hpp"
+#include <cmath>
+#include <cstring>
+#include <cstdint>
+#include <algorithm>
+
+namespace mamba_cpu {
+
+// ---------------------------------------------------------------------------
+// causal_conv1d_update — templated for native BF16/FP32
+//
+// state_ptr may point to a NON-CONTIGUOUS paged KV cache tensor.
+// Explicit strides are passed so the kernel writes directly into the
+// correct memory locations without making a contiguous copy of the full
+// paged tensor (which was the source of the 34-41% direct_copy_kernel).
+//
+//   stride_s_slot  = state.stride(0)  — between cache slots
+//   stride_s_dim   = state.stride(1)  — between conv_dim channels
+//   stride_s_state = state.stride(2)  — between state elements
+//
+// When stride_s_state == 1 (contiguous), the memmove fast path is used.
+// ---------------------------------------------------------------------------
+template <typename scalar_t>
+inline void causal_conv1d_update_kernel(
+    const scalar_t* __restrict__ x_ptr, scalar_t* __restrict__ state_ptr,
+    int64_t stride_s_slot, int64_t stride_s_dim, int64_t stride_s_state,
+    const scalar_t* __restrict__ weight_ptr, const float* __restrict__ bias_ptr,
+    scalar_t* __restrict__ out_ptr, const int32_t* __restrict__ cache_idxs,
+    int32_t pad_slot_id, int64_t batch, int64_t dim, int64_t seqlen,
+    int64_t width, int64_t state_len, bool do_silu) {
+#pragma omp parallel for
+  for (int64_t b = 0; b < batch; ++b) {
+    int64_t cache_idx = (cache_idxs != nullptr) ? cache_idxs[b] : b;
+    if (cache_idx == pad_slot_id) continue;
+
+    for (int64_t t = 0; t < seqlen; ++t) {
+      const scalar_t* x_b = x_ptr + (b * dim * seqlen + t);
+      scalar_t* out_b = out_ptr + (b * dim * seqlen + t);
+      // Base of this slot in the (possibly non-contiguous) paged state
+      scalar_t* s_base = state_ptr + cache_idx * stride_s_slot;
+
+      for (int64_t d = 0; d < dim; ++d) {
+        float x_val = static_cast<float>(x_b[d * seqlen]);
+        scalar_t* sd = s_base + d * stride_s_dim;  // start of this dim's state
+        const scalar_t* w = weight_ptr + d * width;
+
+        // Accumulate in float32 for precision
+        float acc = (bias_ptr != nullptr) ? bias_ptr[d] : 0.0f;
+        for (int64_t k = 0; k < state_len; ++k) {
+          acc += static_cast<float>(w[k]) *
+                 static_cast<float>(sd[k * stride_s_state]);
+        }
+        acc += static_cast<float>(w[state_len]) * x_val;
+
+        // Shift state left and append new input.
+        // Use memmove when contiguous (stride==1); element loop otherwise.
+        if (stride_s_state == 1) {
+          if (state_len > 1)
+            std::memmove(sd, sd + 1, (state_len - 1) * sizeof(scalar_t));
+          if (state_len > 0) sd[state_len - 1] = static_cast<scalar_t>(x_val);
+        } else {
+          for (int64_t k = 0; k < state_len - 1; ++k)
+            sd[k * stride_s_state] = sd[(k + 1) * stride_s_state];
+          if (state_len > 0)
+            sd[(state_len - 1) * stride_s_state] = static_cast<scalar_t>(x_val);
+        }
+
+        if (do_silu) {
+          float sigmoid = (acc >= 0) ? 1.0f / (1.0f + std::exp(-acc))
+                                     : std::exp(acc) / (1.0f + std::exp(acc));
+          acc *= sigmoid;
+        }
+        out_b[d * seqlen] = static_cast<scalar_t>(acc);
+      }
+    }
+  }
+}
+
+// ---------------------------------------------------------------------------
+// selective_state_update
+//
+// Template parameters:
+//   state_t  - dtype of ssm_state cache (typically BFloat16)
+//   input_t  - dtype of x, B, C (typically BFloat16)
+//   out_t    - dtype of output tensor (typically BFloat16)
+//              Write directly — no float32 intermediate buffer needed.
+//
+// A, D, dt_bias are accepted as const float* (they are always float32
+// model parameters in Mamba2).  This eliminates the per-call float32→BF16
+// conversion and the .contiguous() materialisation of the broadcast-expand.
+//
+// dt is accepted as a (N, nheads) scalar-per-head tensor, not as the
+// (N, nheads, head_dim) expansion, so no .contiguous() copy is needed.
+// ---------------------------------------------------------------------------
+template <typename state_t, typename input_t, typename out_t = float>
+inline void selective_state_update_kernel(
+    state_t* __restrict__ state_ptr, int64_t stride_state_n,
+    int64_t stride_state_h, int64_t stride_state_d,
+    const input_t* __restrict__ x_ptr, int64_t stride_x_n, int64_t stride_x_h,
+    // dt: (N, nheads) — scalar per head, NOT expanded to head_dim
+    const float* __restrict__ dt_ptr, int64_t stride_dt_n,
+    // A: (nheads,) float32 — scalar per head
+    const float* __restrict__ A_ptr, const input_t* __restrict__ B_ptr,
+    const input_t* __restrict__ C_ptr, int64_t stride_BC_n, int64_t stride_BC_g,
+    // D: (nheads,) float32 — scalar per head (nullptr if not used)
+    const float* __restrict__ D_ptr,
+    // z: same shape as x (optional)
+    const input_t* __restrict__ z_ptr,
+    // dt_bias: (nheads,) float32 — scalar per head (nullptr if not used)
+    const float* __restrict__ dt_bias_ptr, out_t* __restrict__ out_ptr,
+    int64_t stride_out_n, int64_t stride_out_h,
+    const int32_t* __restrict__ state_batch_indices,
+    const int32_t* __restrict__ dst_state_batch_indices, int32_t null_block_id,
+    const int32_t* __restrict__ num_accepted_tokens,
+    const int32_t* __restrict__ cu_seqlens, int64_t N, int64_t nheads,
+    int64_t ngroups, int64_t dim, int64_t dstate, bool dt_softplus) {
+  using state_vec_t = vec_op::vec_t<state_t>;
+  using input_vec_t = vec_op::vec_t<input_t>;
+  constexpr int VEC_ELEM_NUM = 8;
+
+  int64_t nheads_per_group = nheads / ngroups;
+
+  for (int64_t seq_idx = 0; seq_idx < N; ++seq_idx) {
+    int64_t bos, seq_len;
+    if (cu_seqlens != nullptr) {
+      bos = cu_seqlens[seq_idx];
+      seq_len = cu_seqlens[seq_idx + 1] - bos;
+    } else {
+      bos = seq_idx;
+      seq_len = 1;
+    }
+
+    int64_t state_read_idx = (state_batch_indices != nullptr)
+                                 ? state_batch_indices[seq_idx]
+                                 : seq_idx;
+    if (state_read_idx == null_block_id) continue;
+
+    int64_t state_write_idx = (num_accepted_tokens == nullptr)
+                                  ? ((dst_state_batch_indices != nullptr)
+                                         ? dst_state_batch_indices[seq_idx]
+                                         : state_read_idx)
+                                  : -1;
+
+    state_t* s = state_ptr + state_read_idx * stride_state_n;
+
+    for (int64_t t = 0; t < seq_len; ++t) {
+      int64_t token_idx = bos + t;
+      const input_t* x_tok = x_ptr + token_idx * stride_x_n;
+      // dt: (N, nheads) — one float per head per token
+      const float* dt_tok = dt_ptr + token_idx * stride_dt_n;
+      const input_t* B_tok = B_ptr + token_idx * stride_BC_n;
+      const input_t* C_tok = C_ptr + token_idx * stride_BC_n;
+      out_t* out_tok = out_ptr + token_idx * stride_out_n;
+
+#pragma omp parallel for
+      for (int64_t h = 0; h < nheads; ++h) {
+        int64_t g = h / nheads_per_group;
+        const input_t* x_h = x_tok + h * stride_x_h;
+        const input_t* B_g = B_tok + g * stride_BC_g;
+        const input_t* C_g = C_tok + g * stride_BC_g;
+        out_t* out_h = out_tok + h * stride_out_h;
+        state_t* s_h = s + h * stride_state_h;
+
+        // Read scalars-per-head (A, dt, dt_bias, D) — no per-dim indexing
+        float dt_val = dt_tok[h];
+        if (dt_bias_ptr != nullptr) dt_val += dt_bias_ptr[h];
+        if (dt_softplus) {
+          dt_val = (dt_val <= 20.0f) ? std::log1p(std::exp(dt_val)) : dt_val;
+        }
+        const float A_val = A_ptr[h];  // scalar: same for all dim, dstate
+        const float D_val = (D_ptr != nullptr) ? D_ptr[h] : 0.0f;
+
+        const input_t* z_h =
+            (z_ptr != nullptr) ? z_ptr + token_idx * stride_x_n + h * stride_x_h
+                               : nullptr;
+
+        vec_op::FP32Vec8 dt_vec(dt_val);
+        // dA = exp(A * dt): A and dt are SCALARS per head, so compute once
+        // and broadcast.  This saves 7 redundant std::exp() calls that
+        // FP32Vec8::exp() would otherwise make on the broadcast vector.
+        const float dA_scalar = std::exp(A_val * dt_val);
+        vec_op::FP32Vec8 dA(dA_scalar);  // broadcast
+
+        for (int64_t d = 0; d < dim; ++d) {
+          float x_val = static_cast<float>(x_h[d]);
+
+          vec_op::FP32Vec8 out_vec(0.0f);
+          state_t* s_hd = s_h + d * stride_state_d;
+          const input_t* B_g_base = B_g;
+          const input_t* C_g_base = C_g;
+
+          vec_op::FP32Vec8 x_vec(x_val);
+          // dBx = B * x * dt — same dA for all dstate (A is scalar)
+          // s_new = s * dA + B * x * dt
+
+          int64_t n = 0;
+          for (; n <= dstate - VEC_ELEM_NUM; n += VEC_ELEM_NUM) {
+            vec_op::FP32Vec8 B_v((input_vec_t(B_g_base + n)));
+            vec_op::FP32Vec8 C_v((input_vec_t(C_g_base + n)));
+            vec_op::FP32Vec8 s_v((state_vec_t(s_hd + n)));
+
+            vec_op::FP32Vec8 dBx = B_v * x_vec * dt_vec;
+            vec_op::FP32Vec8 s_new = s_v * dA + dBx;
+
+            state_vec_t(s_new).save(s_hd + n);
+            out_vec = out_vec + s_new * C_v;
+          }
+
+          float out_val = out_vec.reduce_sum();
+          for (; n < dstate; ++n) {
+            // Reuse dA_scalar computed once per head — no exp() re-call
+            float dBx = static_cast<float>(B_g[n]) * x_val * dt_val;
+            float s_new = static_cast<float>(s_hd[n]) * dA_scalar + dBx;
+            s_hd[n] = static_cast<state_t>(s_new);
+            out_val += s_new * static_cast<float>(C_g[n]);
+          }
+
+          if (D_ptr != nullptr) out_val += x_val * D_val;
+          if (z_h != nullptr) {
+            float z_val = static_cast<float>(z_h[d]);
+            float sigmoid = (z_val >= 0)
+                                ? 1.0f / (1.0f + std::exp(-z_val))
+                                : std::exp(z_val) / (1.0f + std::exp(z_val));
+            out_val *= z_val * sigmoid;
+          }
+          out_h[d] = static_cast<out_t>(out_val);
+        }
+      }
+
+      if (num_accepted_tokens != nullptr &&
+          dst_state_batch_indices != nullptr) {
+        int64_t token_dst_idx = dst_state_batch_indices[seq_idx * seq_len + t];
+        if (token_dst_idx != null_block_id && token_dst_idx != state_read_idx) {
+          state_t* dst_s = state_ptr + token_dst_idx * stride_state_n;
+          std::memmove(dst_s, s, nheads * stride_state_h * sizeof(state_t));
+        }
+      }
+    }
+
+    if (num_accepted_tokens == nullptr && state_write_idx != null_block_id &&
+        state_write_idx != state_read_idx) {
+      state_t* dst_s = state_ptr + state_write_idx * stride_state_n;
+      std::memmove(dst_s, s, nheads * stride_state_h * sizeof(state_t));
+    }
+  }
+}
+
+// ---------------------------------------------------------------------------
+// mamba_chunk_scan_fwd
+//
+// Prefill SSM recurrence for Mamba2 / SSD models.
+//
+// Key difference from selective_state_update_kernel (decode path):
+//   - #pragma omp parallel for collapse(2) is OUTSIDE the time loop.
+//     Each thread owns a (batch, head) slice and runs the entire token
+//     sequence without any per-token OpenMP synchronisation overhead.
+//     For seqlen=256, this eliminates 256 thread-barrier launches per batch.
+//
+// `dt` arrives already processed (float32, after bias + softplus + clamp)
+// to keep this kernel simple. Preprocessing is done in the Python wrapper.
+//
+// `states_ptr` points to the [batch, nheads, headdim, dstate] float32 output
+// tensor, pre-initialised by the caller (zero or from initial_states).
+// Each (b, h) slice is private to exactly one thread via collapse(2), so
+// there are no write conflicts.
+//
+// D is treated as a scalar per head ([nheads] float32).
+// ---------------------------------------------------------------------------
+template <typename input_t>
+inline void mamba_chunk_scan_fwd_kernel(
+    float* __restrict__ states_ptr,     // [batch, nheads, headdim, dstate] f32
+    const input_t* __restrict__ x_ptr,  // [seqlen, nheads, headdim]
+    const float* __restrict__ dt_ptr,   // [seqlen, nheads] f32 (preprocessed)
+    const float* __restrict__ A_ptr,    // [nheads] f32
+    const input_t* __restrict__ B_ptr,  // [seqlen, ngroups, dstate]
+    const input_t* __restrict__ C_ptr,  // [seqlen, ngroups, dstate]
+    const float* __restrict__ D_ptr,    // [nheads] f32 (nullable)
+    const input_t* __restrict__ z_ptr,  // [seqlen, nheads, headdim] (nullable)
+    input_t* __restrict__ out_ptr,      // [seqlen, nheads, headdim]
+    const int32_t* __restrict__ cu_seqlens,  // [batch+1] int32
+    int64_t batch, int64_t nheads, int64_t ngroups, int64_t headdim,
+    int64_t dstate) {
+  using input_vec_t = vec_op::vec_t<input_t>;
+  constexpr int VEC_ELEM_NUM = 8;
+
+  const int64_t nheads_per_group = nheads / ngroups;
+  // states layout: [batch, nheads, headdim, dstate] contiguous (caller
+  // guarantee)
+  const int64_t stride_s_b = nheads * headdim * dstate;
+  const int64_t stride_s_h = headdim * dstate;
+  // stride_s_d = dstate, stride_s_n = 1
+
+#pragma omp parallel for collapse(2) schedule(static)
+  for (int64_t b = 0; b < batch; ++b) {
+    for (int64_t h = 0; h < nheads; ++h) {
+      const int64_t seq_start = cu_seqlens[b];
+      const int64_t seq_end = cu_seqlens[b + 1];
+      const int64_t g = h / nheads_per_group;
+
+      const float A_val = A_ptr[h];
+      const float D_val = (D_ptr != nullptr) ? D_ptr[h] : 0.0f;
+
+      // Working state slice: states[b, h, :, :] — float32, headdim * dstate.
+      // Fits in L1/L2 for typical dims (e.g. 64*128*4 = 32 KB).
+      float* s_bh = states_ptr + b * stride_s_b + h * stride_s_h;
+
+      for (int64_t t = seq_start; t < seq_end; ++t) {
+        const input_t* x_h = x_ptr + t * nheads * headdim + h * headdim;
+        const float* dt_h = dt_ptr + t * nheads + h;
+        const input_t* B_g = B_ptr + t * ngroups * dstate + g * dstate;
+        const input_t* C_g = C_ptr + t * ngroups * dstate + g * dstate;
+        const input_t* z_h = (z_ptr != nullptr)
+                                 ? z_ptr + t * nheads * headdim + h * headdim
+                                 : nullptr;
+        input_t* out_h = out_ptr + t * nheads * headdim + h * headdim;
+
+        const float dt_val = *dt_h;
+        const float dA_val = std::exp(A_val * dt_val);
+        const vec_op::FP32Vec8 dA_vec(dA_val);  // broadcast scalar
+        const vec_op::FP32Vec8 dt_vec(dt_val);
+
+        for (int64_t d = 0; d < headdim; ++d) {
+          const float x_val = static_cast<float>(x_h[d]);
+          float* s_bhd = s_bh + d * dstate;  // [dstate] contiguous float32
+
+          // Vectorised SSM update + readout over dstate:
+          //   s_new = s * dA + x * dt * B
+          //   y    += s_new * C
+          int64_t n = 0;
+          vec_op::FP32Vec8 y_vec(0.0f);
+          const vec_op::FP32Vec8 x_vec(x_val);
+
+          for (; n <= dstate - VEC_ELEM_NUM; n += VEC_ELEM_NUM) {
+            const vec_op::FP32Vec8 B_v((input_vec_t(B_g + n)));
+            const vec_op::FP32Vec8 C_v((input_vec_t(C_g + n)));
+            const vec_op::FP32Vec8 s_v(s_bhd + n);
+
+            const vec_op::FP32Vec8 s_new = s_v * dA_vec + x_vec * dt_vec * B_v;
+            s_new.save(s_bhd + n);
+            y_vec = y_vec + s_new * C_v;
+          }
+
+          float y_val = y_vec.reduce_sum();
+
+          // Scalar tail for remaining dstate elements
+          for (; n < dstate; ++n) {
+            const float B_n = static_cast<float>(B_g[n]);
+            const float C_n = static_cast<float>(C_g[n]);
+            const float s_new = s_bhd[n] * dA_val + x_val * dt_val * B_n;
+            s_bhd[n] = s_new;
+            y_val += s_new * C_n;
+          }
+
+          // D skip connection (scalar per head)
+          if (D_ptr != nullptr) y_val += x_val * D_val;
+
+          // z gating: out = y * z * sigmoid(z)  (SiLU)
+          if (z_h != nullptr) {
+            const float z_val = static_cast<float>(z_h[d]);
+            const float sigmoid =
+                (z_val >= 0.0f) ? 1.0f / (1.0f + std::exp(-z_val))
+                                : std::exp(z_val) / (1.0f + std::exp(z_val));
+            y_val *= z_val * sigmoid;
+          }
+
+          out_h[d] = static_cast<input_t>(y_val);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace mamba_cpu

csrc/cpu/torch_bindings.cpp

@@ -213,6 +213,32 @@ void compute_slot_mapping_kernel_impl(const torch::Tensor query_start_loc,
                                       torch::Tensor slot_mapping,
                                       const int64_t block_size);
 
+at::Tensor causal_conv1d_update_cpu_impl(
+    at::Tensor& x, at::Tensor& conv_state, const at::Tensor& weight,
+    const c10::optional<at::Tensor>& bias,
+    const c10::optional<std::string>& activation,
+    const c10::optional<at::Tensor>& conv_state_indices,
+    const c10::optional<at::Tensor>& query_start_loc, int64_t pad_slot_id);
+
+void selective_state_update_cpu_impl(
+    at::Tensor& state, const at::Tensor& x, const at::Tensor& dt,
+    const at::Tensor& A, const at::Tensor& B, const at::Tensor& C,
+    const c10::optional<at::Tensor>& D, const c10::optional<at::Tensor>& z,
+    const c10::optional<at::Tensor>& dt_bias, bool dt_softplus,
+    const c10::optional<at::Tensor>& state_batch_indices,
+    const c10::optional<at::Tensor>& dst_state_batch_indices,
+    int64_t null_block_id, at::Tensor& out,
+    const c10::optional<at::Tensor>& num_accepted_tokens,
+    const c10::optional<at::Tensor>& cu_seqlens);
+
+void mamba_chunk_scan_fwd_cpu_impl(at::Tensor& out, at::Tensor& final_states,
+                                   const at::Tensor& x, const at::Tensor& dt,
+                                   const at::Tensor& A, const at::Tensor& B,
+                                   const at::Tensor& C,
+                                   const c10::optional<at::Tensor>& D,
+                                   const c10::optional<at::Tensor>& z,
+                                   const at::Tensor& cu_seqlens);
+
 void init_cpu_memory_env(std::vector<int64_t> node_ids);
 
 namespace cpu_utils {
@@ -595,6 +621,30 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "block_size) -> ()",
       &compute_slot_mapping_kernel_impl);
 
+  // Mamba CPU kernels
+  ops.def(
+      "causal_conv1d_update_cpu_vec("
+      "Tensor(a0!) x, Tensor(a1!) conv_state, Tensor weight, "
+      "Tensor? bias, str? activation, Tensor? conv_state_indices, "
+      "Tensor? query_start_loc, SymInt pad_slot_id) -> Tensor",
+      &causal_conv1d_update_cpu_impl);
+
+  ops.def(
+      "selective_state_update_cpu("
+      "Tensor(a0!) state, Tensor x, Tensor dt, Tensor A, Tensor B, Tensor C, "
+      "Tensor? D, Tensor? z, Tensor? dt_bias, bool dt_softplus, "
+      "Tensor? state_batch_indices, Tensor? dst_state_batch_indices, "
+      "SymInt null_block_id, Tensor(a13!) out, "
+      "Tensor? num_accepted_tokens, Tensor? cu_seqlens) -> ()",
+      &selective_state_update_cpu_impl);
+
+  ops.def(
+      "mamba_chunk_scan_fwd_cpu("
+      "Tensor(a0!) out, Tensor(a1!) final_states, "
+      "Tensor x, Tensor dt, Tensor A, Tensor B, Tensor C, "
+      "Tensor? D, Tensor? z, Tensor cu_seqlens) -> ()",
+      &mamba_chunk_scan_fwd_cpu_impl);
+
   ops.def("init_cpu_memory_env(SymInt[] node_ids) -> ()", &init_cpu_memory_env);
 
   // Speculative decoding kernels

csrc/libtorch_stable/quantization/marlin/marlin_int4_fp8_preprocess.cu

@@ -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()),

docs/design/attention_backends.md

@@ -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 |

docs/design/fusions.md

@@ -306,7 +306,7 @@ Supported quantization scheme/hardware combinations:
 
 - Pass: [`vllm/compilation/passes/fusion/rms_quant_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/rms_quant_fusion.py)
 - ROCm AITER pass: [`vllm/compilation/passes/fusion/rocm_aiter_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/rocm_aiter_fusion.py)
-- CUDA/HIP kernels: [`csrc/layernorm_quant_kernels.cu`](https://github.com/vllm-project/vllm/blob/main/csrc/layernorm_quant_kernels.cu)
+- CUDA/HIP kernels: [`csrc/libtorch_stable/layernorm_quant_kernels.cu`](https://github.com/vllm-project/vllm/blob/main/csrc/libtorch_stable/layernorm_quant_kernels.cu)
 
 ### SiLU+Mul + Quantization (`fuse_act_quant`)
 
@@ -332,7 +332,7 @@ Supported quantization scheme/hardware combinations:
 - Pass: [`vllm/compilation/passes/fusion/act_quant_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/act_quant_fusion.py)
 - ROCm AITER pass: [`vllm/compilation/passes/fusion/rocm_aiter_fusion.py`](https://github.com/vllm-project/vllm/blob/main/vllm/compilation/passes/fusion/rocm_aiter_fusion.py)
 - CUDA/HIP kernels: [`csrc/quantization/`](https://github.com/vllm-project/vllm/blob/main/csrc/quantization/)
-- Fused SiLU+Mul+BlockQuant kernel: [`csrc/quantization/fused_kernels/fused_silu_mul_block_quant.cu`](https://github.com/vllm-project/vllm/blob/main/csrc/quantization/fused_kernels/fused_silu_mul_block_quant.cu)
+- Fused SiLU+Mul+BlockQuant kernel: [`csrc/libtorch_stable/quantization/fused_kernels/fused_silu_mul_block_quant.cu`](https://github.com/vllm-project/vllm/blob/main/csrc/libtorch_stable/quantization/fused_kernels/fused_silu_mul_block_quant.cu)
 
 ### RMSNorm + Padding (`fuse_act_padding`)
 

docs/features/kv_offloading_usage.md

@@ -68,13 +68,14 @@ vllm serve <model> \
 | --- | --- | --- | --- | --- |
 | `spec_name` | no | `CPUOffloadingSpec` | both | Set to `TieringOffloadingSpec` for multi-tier. |
 | `cpu_bytes_to_use` | yes | — | both | Total bytes of host memory reserved for the CPU tier across all workers (not per-worker). |
-| `block_size` | no | GPU block size | both | Offloaded block size in tokens; must be a multiple of the GPU block size. |
+| `block_size` | no | GPU block size | both | Offloaded block size in tokens; must be a multiple of the GPU block size. Mutually exclusive with `blocks_per_chunk`. |
+| `blocks_per_chunk` | no | `1` | both | Offloaded chunk size in GPU blocks; must be > 0. Alternative to `block_size` for models whose KV cache groups have different block sizes. |
 | `eviction_policy` | no | `lru` | both | Primary tier policy: `lru` or `arc`. |
 | `store_threshold` | no | `0` | single-tier | Min lookups before a block is offloaded. Values ≥ 2 are rejected by `TieringOffloadingSpec`. |
 | `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), 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
@@ -179,7 +180,7 @@ Rather than embedding `host`/`port` in each `secondary_tiers` entry, set them on
 
 - `cpu_bytes_to_use`: a bigger CPU tier means fewer trips to slower secondary tiers and a higher hit rate. The value is total across all workers, not per-worker. Leave headroom for the rest of the host workload.
 - For single-tier (CPU-only) setups, set `cpu_bytes_to_use` larger than the aggregate GPU KV cache. Because offloading is immediate, a smaller CPU tier just mirrors what the GPU already holds and adds no hit rate.
-- `block_size`: larger offloaded blocks reduce per-block bookkeeping overhead but increase the granularity of lookups. Must be a multiple of the GPU block size.
+- `block_size` / `blocks_per_chunk`: larger offloaded chunks reduce per-block bookkeeping overhead but increase the granularity of lookups.
 - FS thread counts: tune `n_read_threads` and `n_write_threads` to the parallelism your storage can sustain. Reads are latency-sensitive on the prefill path, so prefer more read threads when prefill hit rates are high.
 - Sharing `root_dir` across runs: runs with the same model, `block_size`, parallelism layout, and dtype share files under the same `<digest>` subdirectory. Changing any of these produces a new subdirectory; old ones are orphaned but harmless. Delete them to reclaim disk.
 

docs/features/nixl_connector_compatibility.md

@@ -59,7 +59,7 @@ th:not(:first-child) {
 
 <sup>1</sup> P and D instances must use the same speculation configuration.
 
-<sup>2</sup> Requires `FLASH_ATTN` or `FLASHINFER` backend **and** `HND` KV cache layout. Enable via `--kv-transfer-config '{"kv_connector_extra_config": {"enable_cross_layers_blocks": "True"}}'`.
+<sup>2</sup> Cross-layer contiguity is achieved by using a `BLHNC` layout (set via `VLLM_KV_CACHE_LAYOUT=BLHNC` or `--enable-cross-layers`).
 
 <sup>3</sup> Supported only when HMA is **not** required (i.e., non-hybrid models). Block IDs are remapped automatically. Only P block size < D block size is supported.
 

docs/features/nixl_connector_usage.md

@@ -414,15 +414,6 @@ Support use case: Prefill with 'HND' and decode with 'NHD' with experimental con
 --kv-transfer-config '{..., "enable_permute_local_kv":"True"}'
 ```
 
-### Cross layers blocks
-
-By default, this feature is disabled. On attention backends that support this feature, each logical block is contiguous in physical memory. This reduces the number of buffers that need to be transferred.
-To enable this feature:
-
-```bash
---kv-transfer-config '{..., "kv_connector_extra_config": {"enable_cross_layers_blocks": "True"}}'
-```
-
 ## Metrics Reference
 
 vLLM periodically logs a `KV Transfer metrics` line summarising NIXL transfer

docs/getting_started/installation/gpu.xpu.inc.md

@@ -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]

docs/getting_started/quickstart.md

@@ -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.

docs/models/hardware_supported_models/xpu.md

@@ -31,10 +31,8 @@
 | THUDM/CodeGeex4-All-9B                             | CodeGeexForCausalLM                              | ✅                    |                        |       |
 | chuhac/TeleChat2-35B                               | LlamaForCausalLM (TeleChat2 based on Llama arch) | ✅                    |                        |       |
 | 01-ai/Yi1.5-34B-Chat                               | YiForCausalLM                                    | ✅                    |                        |       |
-| THUDM/CodeGeex4-All-9B                             | CodeGeexForCausalLM                              | ✅                    |                        |       |
 | deepseek-ai/DeepSeek-Coder-33B-base                | DeepSeekCoderForCausalLM                         | ✅                    |                        |       |
 | meta-llama/Llama-2-13b-chat-hf                     | LlamaForCausalLM                                 | ✅                    |                        |       |
-| THUDM/CodeGeex4-All-9B                             | CodeGeexForCausalLM                              | ✅                    |                        |       |
 | Qwen/Qwen1.5-14B-Chat                              | QwenForCausalLM                                  | ✅                    |                        |       |
 | Qwen/Qwen1.5-32B-Chat                              | QwenForCausalLM                                  | ✅                    |                        |       |
 | RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8-dynamic    | LlamaForCausalLM                                 |                       | ✅                     |       |

docs/serving/online_serving/speech_to_text.md

@@ -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:**
 

rust/Cargo.lock

@@ -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",
 ]
@@ -5560,6 +5576,7 @@ dependencies = [
  "tracing",
  "tracing-subscriber",
  "uuid",
+ "vllm-bench",
  "vllm-chat",
  "vllm-engine-core-client",
  "vllm-managed-engine",
@@ -5729,6 +5746,7 @@ dependencies = [
  "tokio-stream",
  "tokio-util",
  "tonic",
+ "tonic-health",
  "tonic-prost",
  "tonic-prost-build",
  "tower",
@@ -6320,9 +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",

rust/Cargo.toml

@@ -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"] }
@@ -132,6 +133,7 @@ trait-set = "0.3.0"
 url = "2.5.7"
 uuid = { version = "1.22.0", features = ["v4"] }
 validator = { version = "0.20.0", features = ["derive"] }
+vllm-bench = { path = "src/bench" }
 vllm-chat = { path = "src/chat" }
 vllm-engine-core-client = { path = "src/engine-core-client" }
 vllm-llm = { path = "src/llm" }
@@ -142,7 +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",

rust/src/bench/Cargo.toml

@@ -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
 

rust/src/bench/src/backends/pooling.rs

@@ -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!({

rust/src/bench/src/benchmark.rs

@@ -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);
         }

rust/src/bench/src/cli.rs

@@ -3,8 +3,6 @@
 
 use std::fmt;
 
-use clap::Parser;
-
 /// Backend type for the benchmark endpoint.
 #[derive(clap::ValueEnum, Debug, Clone, Copy, PartialEq, Eq)]
 pub enum BackendKind {
@@ -77,7 +75,7 @@ pub enum DatasetName {
     ShareGpt,
     #[value(name = "sonnet")]
     Sonnet,
-    #[value(name = "speed-bench")]
+    #[value(name = "speed-bench", alias = "speed_bench")]
     SpeedBench,
     #[value(name = "hf")]
     Hf,
@@ -144,13 +142,8 @@ impl fmt::Display for SpeedBenchConfig {
 }
 
 /// High-performance benchmark client for vLLM serving endpoints.
-#[derive(Parser, Debug, Clone)]
-#[command(
-    name = "vllm-bench",
-    about = "Benchmark online serving throughput",
-    version
-)]
-pub struct Cli {
+#[derive(clap::Args, Debug, Clone)]
+pub struct BenchServeArgs {
     /// The type of backend or endpoint to use for the benchmark.
     #[arg(long, default_value = "openai")]
     pub backend: BackendKind,
@@ -659,7 +652,7 @@ pub struct Cli {
     pub lora_assignment: LoraAssignment,
 }
 
-impl Cli {
+impl BenchServeArgs {
     /// Resolve the base URL from explicit --base-url or from --host/--port.
     pub fn resolve_base_url(&self) -> String {
         if let Some(ref base) = self.base_url {

rust/src/bench/src/config.rs

@@ -4,7 +4,9 @@
 use std::collections::HashMap;
 use std::sync::Arc;
 
-use crate::cli::{BackendKind, Cli, DatasetName, LoraAssignment, RampUpStrategy, SpeedBenchConfig};
+use crate::cli::{
+    BackendKind, BenchServeArgs, DatasetName, LoraAssignment, RampUpStrategy, SpeedBenchConfig,
+};
 use crate::datasets::random_mm::{MmBucketKey, MmLimitPerPrompt};
 use crate::error::{BenchError, Result};
 
@@ -215,63 +217,63 @@ pub struct BenchConfig {
 }
 
 impl BenchConfig {
-    pub fn from_cli(cli: &Cli) -> Result<Self> {
-        if cli.burstiness <= 0.0 {
+    pub fn from_args(args: &BenchServeArgs) -> Result<Self> {
+        if args.burstiness <= 0.0 {
             return Err(BenchError::Config("Burstiness must be positive".into()));
         }
 
-        if cli.num_prompts == 0 {
+        if args.num_prompts == 0 {
             return Err(BenchError::Config(
                 "--num-prompts must be at least 1".into(),
             ));
         }
 
-        if cli.request_rate <= 0.0 && !cli.request_rate.is_infinite() {
+        if args.request_rate <= 0.0 && !args.request_rate.is_infinite() {
             return Err(BenchError::Config(
                 "--request-rate must be positive (or inf)".into(),
             ));
         }
-        if cli.max_model_len == Some(0) {
+        if args.max_model_len == Some(0) {
             return Err(BenchError::Config(
                 "--max-model-len must be at least 1".into(),
             ));
         }
 
-        let base_url = cli.resolve_base_url();
-        let api_url = cli.resolve_api_url();
+        let base_url = args.resolve_base_url();
+        let api_url = args.resolve_api_url();
 
-        let extra_headers = cli.parse_headers()?;
-        let mut extra_body = cli.parse_extra_body()?;
+        let extra_headers = args.parse_headers()?;
+        let mut extra_body = args.parse_extra_body()?;
 
         // Merge sampling parameters into extra_body (matches Python behavior).
         // Python collects non-None sampling params and merges them UNDER extra_body,
         // meaning extra_body keys take precedence over sampling params.
         {
             let mut sampling_params = serde_json::Map::new();
-            if let Some(v) = cli.top_p {
+            if let Some(v) = args.top_p {
                 sampling_params.insert("top_p".into(), serde_json::json!(v));
             }
-            if let Some(v) = cli.top_k {
+            if let Some(v) = args.top_k {
                 sampling_params.insert("top_k".into(), serde_json::json!(v));
             }
-            if let Some(v) = cli.min_p {
+            if let Some(v) = args.min_p {
                 sampling_params.insert("min_p".into(), serde_json::json!(v));
             }
-            if let Some(v) = cli.temperature {
+            if let Some(v) = args.temperature {
                 sampling_params.insert("temperature".into(), serde_json::json!(v));
             }
-            if let Some(v) = cli.frequency_penalty {
+            if let Some(v) = args.frequency_penalty {
                 sampling_params.insert("frequency_penalty".into(), serde_json::json!(v));
             }
-            if let Some(v) = cli.presence_penalty {
+            if let Some(v) = args.presence_penalty {
                 sampling_params.insert("presence_penalty".into(), serde_json::json!(v));
             }
-            if let Some(v) = cli.repetition_penalty {
+            if let Some(v) = args.repetition_penalty {
                 sampling_params.insert("repetition_penalty".into(), serde_json::json!(v));
             }
 
             if !sampling_params.is_empty() {
-                if !cli.backend.is_openai_compatible() {
+                if !args.backend.is_openai_compatible() {
                     return Err(BenchError::Config(
                         "Sampling parameters are only supported by openai-compatible backends."
                             .into(),
@@ -286,10 +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,
@@ -299,7 +309,7 @@ impl BenchConfig {
         }
 
         // Parse metadata
-        let metadata = match &cli.metadata {
+        let metadata = match &args.metadata {
             None => None,
             Some(items) => {
                 let mut pairs = Vec::new();
@@ -314,24 +324,24 @@ impl BenchConfig {
         };
 
         // Parse goodput SLOs
-        let goodput = parse_goodput(&cli.goodput)?;
+        let goodput = parse_goodput(&args.goodput)?;
 
         // Parse ramp-up config
-        let ramp_up = parse_ramp_up(cli)?;
+        let ramp_up = parse_ramp_up(args)?;
 
         // Default percentile metrics based on backend type
-        let default_percentile_metrics = if cli.backend.is_pooling() {
+        let default_percentile_metrics = if args.backend.is_pooling() {
             "e2el"
         } else {
             "ttft,tpot,itl,e2el"
         };
         let percentile_metrics_str =
-            cli.percentile_metrics.as_deref().unwrap_or(default_percentile_metrics);
+            args.percentile_metrics.as_deref().unwrap_or(default_percentile_metrics);
         let selected_percentile_metrics: Vec<String> =
             percentile_metrics_str.split(',').map(|s| s.trim().to_string()).collect();
 
-        let metric_percentiles = parse_percentiles(&cli.metric_percentiles, false)?;
-        let sweep_summary_percentiles = cli
+        let metric_percentiles = parse_percentiles(&args.metric_percentiles, false)?;
+        let sweep_summary_percentiles = args
             .sweep_summary_percentiles
             .as_deref()
             .map(|raw| parse_percentiles(raw, true))
@@ -344,38 +354,38 @@ impl BenchConfig {
             selected_percentiles.push(90.0);
         }
 
-        let tokenizer_id = if cli.skip_tokenizer_init {
+        let tokenizer_id = if args.skip_tokenizer_init {
             None
         } else {
-            Some(cli.tokenizer.clone().or_else(|| cli.model.clone()).unwrap_or_default())
+            args.tokenizer.clone().or_else(|| args.model.clone())
         };
 
         // Resolve input/output lengths
-        let random_input_len = cli.resolved_random_input_len();
-        let random_output_len = cli.resolved_random_output_len();
-        let per_turn_input_len = cli.resolved_per_turn_input_len();
+        let random_input_len = args.resolved_random_input_len();
+        let random_output_len = args.resolved_random_output_len();
+        let per_turn_input_len = args.resolved_per_turn_input_len();
 
         // Normalized multi-turn turn counts (computed in validation block below, defaults
         // to num_turns if multi-turn mode is not active)
-        let mut multi_turn_min_turns = cli.multi_turn_num_turns;
-        let mut multi_turn_max_turns = cli.multi_turn_num_turns;
+        let mut multi_turn_min_turns = args.multi_turn_num_turns;
+        let mut multi_turn_max_turns = args.multi_turn_num_turns;
 
         // For random datasets with openai-compatible backends, default to ignore_eos.
         // Exception: multi-turn mode, where ignore_eos causes unbounded context growth
         // across turns. Multi-turn uses min_tokens instead for output length control.
         // Pooling backends don't generate tokens, so ignore_eos is irrelevant.
-        let ignore_eos = if cli.backend.is_pooling() {
+        let ignore_eos = if args.backend.is_pooling() {
             false
         } else {
-            cli.ignore_eos
-                || ((cli.dataset_name == DatasetName::Random
-                    || cli.dataset_name == DatasetName::RandomMm)
-                    && cli.backend.is_openai_compatible()
-                    && !cli.multi_turn)
+            args.ignore_eos
+                || ((args.dataset_name == DatasetName::Random
+                    || args.dataset_name == DatasetName::RandomMm)
+                    && args.backend.is_openai_compatible()
+                    && !args.multi_turn)
         };
 
         // Pooling backends don't support multi-turn
-        if cli.backend.is_pooling() && cli.multi_turn {
+        if args.backend.is_pooling() && args.multi_turn {
             return Err(BenchError::Config(
                 "Pooling/embedding backends do not support --multi-turn".into(),
             ));
@@ -383,7 +393,7 @@ impl BenchConfig {
 
         // LoRA validation. Adapter names must be non-empty after trim; pooling
         // backends are out of scope (vLLM LoRA routing is for generative paths).
-        let lora_modules = match cli.lora_modules.as_ref() {
+        let lora_modules = match args.lora_modules.as_ref() {
             None => None,
             Some(names) => {
                 if names.is_empty() {
@@ -391,7 +401,7 @@ impl BenchConfig {
                         "--lora-modules requires at least one adapter name".into(),
                     ));
                 }
-                if cli.backend.is_pooling() {
+                if args.backend.is_pooling() {
                     return Err(BenchError::Config(
                         "--lora-modules is not supported for pooling/embedding backends".into(),
                     ));
@@ -411,18 +421,18 @@ impl BenchConfig {
         };
 
         // Random-MM validation and config parsing
-        let (random_mm_limit, random_mm_buckets) = if cli.dataset_name == DatasetName::RandomMm {
-            if cli.backend != BackendKind::OpenaiChat {
+        let (random_mm_limit, random_mm_buckets) = if args.dataset_name == DatasetName::RandomMm {
+            if args.backend != BackendKind::OpenaiChat {
                 return Err(BenchError::Config(
                     "Multi-modal content (images) is only supported on 'openai-chat' backend."
                         .into(),
                 ));
             }
             let limit = crate::datasets::random_mm::parse_limit_mm_per_prompt(
-                &cli.random_mm_limit_mm_per_prompt,
+                &args.random_mm_limit_mm_per_prompt,
             )?;
             let buckets =
-                crate::datasets::random_mm::parse_bucket_config(&cli.random_mm_bucket_config)?;
+                crate::datasets::random_mm::parse_bucket_config(&args.random_mm_bucket_config)?;
             (limit, buckets)
         } else {
             (MmLimitPerPrompt::default(), Vec::new())
@@ -432,18 +442,18 @@ impl BenchConfig {
         // sonnet (uses built-in Shakespeare's sonnets).
 
         // Range ratio (Python semantics: [len*(1-r), len*(1+r)], each r in [0,1))
-        let random_range_ratio = RangeRatio::parse(&cli.random_range_ratio)?;
+        let random_range_ratio = RangeRatio::parse(&args.random_range_ratio)?;
 
         // Batched inputs only make sense for pooling backends (the generation
         // backends send one prompt per request).
-        if cli.random_batch_size == 0 {
+        if args.random_batch_size == 0 {
             return Err(BenchError::Config(
                 "--random-batch-size must be at least 1".into(),
             ));
         }
-        if cli.random_batch_size > 1
-            && !cli.backend.is_pooling()
-            && cli.dataset_name != DatasetName::RandomRerank
+        if args.random_batch_size > 1
+            && !args.backend.is_pooling()
+            && args.dataset_name != DatasetName::RandomRerank
         {
             return Err(BenchError::Config(
                 "--random-batch-size > 1 is only supported with embeddings/pooling backends".into(),
@@ -451,16 +461,16 @@ impl BenchConfig {
         }
 
         // random-rerank validation (mirrors Python RandomDatasetForReranking)
-        let is_reranker = !cli.no_reranker;
-        if cli.dataset_name == DatasetName::RandomRerank {
-            if !cli.backend.is_pooling() {
+        let is_reranker = !args.no_reranker;
+        if args.dataset_name == DatasetName::RandomRerank {
+            if !args.backend.is_pooling() {
                 return Err(BenchError::Config(
                     "--dataset-name random-rerank requires an embeddings/pooling backend \
                      (e.g. --backend vllm-rerank)"
                         .into(),
                 ));
             }
-            if !is_reranker && (cli.num_prompts < 2 || cli.random_batch_size < 2) {
+            if !is_reranker && (args.num_prompts < 2 || args.random_batch_size < 2) {
                 return Err(BenchError::Config(
                     "--no-reranker requires --num-prompts > 1 and --random-batch-size > 1 \
                      (the query is folded into the first batch slot)"
@@ -470,8 +480,8 @@ impl BenchConfig {
         }
 
         // Custom dataset validation
-        if cli.dataset_name == DatasetName::Custom {
-            match cli.dataset_path.as_deref() {
+        if args.dataset_name == DatasetName::Custom {
+            match args.dataset_path.as_deref() {
                 None => {
                     return Err(BenchError::Config(
                         "--dataset-path is required for --dataset-name custom \
@@ -486,38 +496,38 @@ impl BenchConfig {
                 }
                 _ => {}
             }
-            if !cli.skip_chat_template {
-                eprintln!(
-                    "NOTE: client-side chat template rendering is not supported; custom \
-                     dataset prompts are sent raw (equivalent to --skip-chat-template)."
+            if !args.skip_chat_template {
+                tracing::warn!(
+                    dataset = "custom",
+                    "client-side chat template rendering is unsupported; sending prompts raw"
                 );
             }
         }
 
         // Prefix repetition validation
-        if cli.dataset_name == DatasetName::PrefixRepetition {
-            if cli.prefix_repetition_num_prefixes == 0 {
+        if args.dataset_name == DatasetName::PrefixRepetition {
+            if args.prefix_repetition_num_prefixes == 0 {
                 return Err(BenchError::Config(
                     "--prefix-repetition-num-prefixes must be at least 1".into(),
                 ));
             }
-            if cli.num_prompts < cli.prefix_repetition_num_prefixes {
+            if args.num_prompts < args.prefix_repetition_num_prefixes {
                 return Err(BenchError::Config(format!(
                     "--num-prompts ({}) must be >= --prefix-repetition-num-prefixes ({})",
-                    cli.num_prompts, cli.prefix_repetition_num_prefixes
+                    args.num_prompts, args.prefix_repetition_num_prefixes
                 )));
             }
         }
 
         // HF dataset validation
-        if cli.dataset_name == DatasetName::Hf && cli.dataset_path.is_none() {
+        if args.dataset_name == DatasetName::Hf && args.dataset_path.is_none() {
             return Err(BenchError::Config(
                 "--dataset-path is required for --dataset-name hf \
                  (set to a HuggingFace dataset ID, e.g. 'allenai/WildChat-4.8M')"
                     .into(),
             ));
         }
-        if let Some(len) = cli.hf_output_len
+        if let Some(len) = args.hf_output_len
             && len == 0
         {
             return Err(BenchError::Config(
@@ -526,13 +536,13 @@ impl BenchConfig {
         }
 
         // Multi-turn validation
-        if cli.multi_turn {
-            if cli.backend != BackendKind::OpenaiChat {
+        if args.multi_turn {
+            if args.backend != BackendKind::OpenaiChat {
                 return Err(BenchError::Config(
                     "--multi-turn requires --backend openai-chat".into(),
                 ));
             }
-            if cli.multi_turn_num_turns == 0 {
+            if args.multi_turn_num_turns == 0 {
                 return Err(BenchError::Config(
                     "--multi-turn-num-turns must be at least 1".into(),
                 ));
@@ -541,18 +551,18 @@ impl BenchConfig {
             // Normalize and validate min/max turns. ShareGPT only consumes max_turns
             // (the loader walks all available turns up to the cap), so the
             // min/num/max coupling used for synthetic generation does not apply.
-            if cli.dataset_name == DatasetName::ShareGpt {
-                if cli.multi_turn_max_turns == 1 {
+            if args.dataset_name == DatasetName::ShareGpt {
+                if args.multi_turn_max_turns == 1 {
                     return Err(BenchError::Config(
                         "--multi-turn-max-turns must be at least 2 for ShareGPT multi-turn".into(),
                     ));
                 }
             } else {
                 (multi_turn_min_turns, multi_turn_max_turns) =
-                    match (cli.multi_turn_min_turns, cli.multi_turn_max_turns) {
-                        (0, 0) => (cli.multi_turn_num_turns, cli.multi_turn_num_turns),
-                        (m, 0) => (m, cli.multi_turn_num_turns),
-                        (0, x) => (cli.multi_turn_num_turns, x),
+                    match (args.multi_turn_min_turns, args.multi_turn_max_turns) {
+                        (0, 0) => (args.multi_turn_num_turns, args.multi_turn_num_turns),
+                        (m, 0) => (m, args.multi_turn_num_turns),
+                        (0, x) => (args.multi_turn_num_turns, x),
                         (m, x) => (m, x),
                     };
                 if multi_turn_min_turns < 1 {
@@ -568,15 +578,16 @@ 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"
                 );
             }
 
             // Validate prefix sharing ratios
-            let pg = cli.multi_turn_prefix_global_ratio;
-            let pc = cli.multi_turn_prefix_conversation_ratio;
+            let pg = args.multi_turn_prefix_global_ratio;
+            let pc = args.multi_turn_prefix_conversation_ratio;
             if !(0.0..=1.0).contains(&pg) {
                 return Err(BenchError::Config(
                     "--multi-turn-prefix-global-ratio must be in [0.0, 1.0]".into(),
@@ -592,20 +603,20 @@ impl BenchConfig {
                     "--multi-turn-prefix-global-ratio + --multi-turn-prefix-conversation-ratio must be < 1.0 (unique suffix required)".into(),
                 ));
             }
-            if (pg > 0.0 || pc > 0.0) && cli.dataset_name != DatasetName::Random {
+            if (pg > 0.0 || pc > 0.0) && args.dataset_name != DatasetName::Random {
                 return Err(BenchError::Config(
                     "Prefix sharing (--multi-turn-prefix-global-ratio / --multi-turn-prefix-conversation-ratio) only works with --dataset-name random".into(),
                 ));
             }
         }
 
-        if !(cli.steady_state_threshold > 0.0 && cli.steady_state_threshold <= 1.0) {
+        if !(args.steady_state_threshold > 0.0 && args.steady_state_threshold <= 1.0) {
             return Err(BenchError::Config(format!(
                 "--steady-state-threshold must be in (0.0, 1.0], got {}",
-                cli.steady_state_threshold
+                args.steady_state_threshold
             )));
         }
-        if let Some(mw) = cli.steady_state_min_window
+        if let Some(mw) = args.steady_state_min_window
             && mw < 0.0
         {
             return Err(BenchError::Config(format!(
@@ -613,122 +624,122 @@ impl BenchConfig {
             )));
         }
 
-        if cli.profile_batch_threshold.is_some() && !cli.profile {
+        if args.profile_batch_threshold.is_some() && !args.profile {
             return Err(BenchError::Config(
                 "--profile-batch-threshold requires --profile".into(),
             ));
         }
-        if cli.profile_duration <= 0.0 {
+        if args.profile_duration <= 0.0 {
             return Err(BenchError::Config(
                 "--profile-duration must be positive".into(),
             ));
         }
-        if cli.profile_batch_threshold.is_none() && cli.profile_duration != 5.0 {
+        if args.profile_batch_threshold.is_none() && args.profile_duration != 5.0 {
             return Err(BenchError::Config(
                 "--profile-duration requires --profile-batch-threshold".into(),
             ));
         }
 
         Ok(BenchConfig {
-            backend: cli.backend,
+            backend: args.backend,
             base_url,
             api_url,
-            model: cli.model.clone(),
-            model_name: cli.served_model_name.clone(),
+            model: args.model.clone(),
+            model_name: args.served_model_name.clone(),
             tokenizer_id,
-            tokenizer_mode: cli.tokenizer_mode.clone(),
-            trust_remote_code: cli.trust_remote_code,
-            skip_tokenizer_init: cli.skip_tokenizer_init,
-            dataset_name: cli.dataset_name,
-            dataset_path: cli.dataset_path.clone(),
-            max_model_len: cli.max_model_len,
+            tokenizer_mode: args.tokenizer_mode.clone(),
+            trust_remote_code: args.trust_remote_code,
+            skip_tokenizer_init: args.skip_tokenizer_init,
+            dataset_name: args.dataset_name,
+            dataset_path: args.dataset_path.clone(),
+            max_model_len: args.max_model_len,
             random_input_len,
             random_output_len,
-            random_prefix_len: cli.random_prefix_len,
+            random_prefix_len: args.random_prefix_len,
             random_range_ratio,
-            random_batch_size: cli.random_batch_size,
+            random_batch_size: args.random_batch_size,
             is_reranker,
-            custom_output_len: cli.output_len.map(|v| v as i64).unwrap_or(cli.custom_output_len),
-            prefix_repetition_prefix_len: cli.prefix_repetition_prefix_len,
-            prefix_repetition_suffix_len: cli.prefix_repetition_suffix_len,
-            prefix_repetition_num_prefixes: cli.prefix_repetition_num_prefixes,
-            prefix_repetition_output_len: cli
+            custom_output_len: args.output_len.map(|v| v as i64).unwrap_or(args.custom_output_len),
+            prefix_repetition_prefix_len: args.prefix_repetition_prefix_len,
+            prefix_repetition_suffix_len: args.prefix_repetition_suffix_len,
+            prefix_repetition_num_prefixes: args.prefix_repetition_num_prefixes,
+            prefix_repetition_output_len: args
                 .output_len
-                .unwrap_or(cli.prefix_repetition_output_len),
-            random_cache_hit_fraction: cli.random_cache_hit_fraction,
-            random_cache_ratio: cli.random_cache_ratio,
-            sharegpt_output_len: cli.sharegpt_output_len,
-            sonnet_input_len: cli.sonnet_input_len,
-            sonnet_output_len: cli.sonnet_output_len,
-            sonnet_prefix_len: cli.sonnet_prefix_len,
-            no_oversample: cli.no_oversample,
-            disable_shuffle: cli.disable_shuffle,
-            num_prompts: cli.num_prompts,
-            request_rate: cli.request_rate,
-            burstiness: cli.burstiness,
-            max_concurrency: cli.max_concurrency,
-            steady_state_threshold: cli.steady_state_threshold,
-            steady_state_min_window: cli.steady_state_min_window,
-            no_steady_state: cli.no_steady_state,
-            disable_tqdm: cli.disable_tqdm,
-            num_warmups: cli.num_warmups,
-            profile: cli.profile,
-            profile_batch_threshold: cli.profile_batch_threshold,
-            profile_duration: cli.profile_duration,
-            save_result: cli.save_result,
-            save_detailed: cli.save_detailed,
-            append_result: cli.append_result,
-            result_dir: cli.result_dir.clone(),
-            result_filename: cli.result_filename.clone(),
-            seed: cli.seed,
+                .unwrap_or(args.prefix_repetition_output_len),
+            random_cache_hit_fraction: args.random_cache_hit_fraction,
+            random_cache_ratio: args.random_cache_ratio,
+            sharegpt_output_len: args.sharegpt_output_len,
+            sonnet_input_len: args.sonnet_input_len,
+            sonnet_output_len: args.sonnet_output_len,
+            sonnet_prefix_len: args.sonnet_prefix_len,
+            no_oversample: args.no_oversample,
+            disable_shuffle: args.disable_shuffle,
+            num_prompts: args.num_prompts,
+            request_rate: args.request_rate,
+            burstiness: args.burstiness,
+            max_concurrency: args.max_concurrency,
+            steady_state_threshold: args.steady_state_threshold,
+            steady_state_min_window: args.steady_state_min_window,
+            no_steady_state: args.no_steady_state,
+            disable_tqdm: args.disable_tqdm,
+            num_warmups: args.num_warmups,
+            profile: args.profile,
+            profile_batch_threshold: args.profile_batch_threshold,
+            profile_duration: args.profile_duration,
+            save_result: args.save_result,
+            save_detailed: args.save_detailed,
+            append_result: args.append_result,
+            result_dir: args.result_dir.clone(),
+            result_filename: args.result_filename.clone(),
+            seed: args.seed,
             ignore_eos,
-            insecure: cli.insecure,
+            insecure: args.insecure,
             selected_percentile_metrics,
             selected_percentiles,
             sweep_summary_percentiles,
-            label: cli.label.clone(),
-            logprobs: cli.logprobs,
-            request_id_prefix: cli.get_request_id_prefix(),
-            ready_check_timeout_sec: cli.ready_check_timeout_sec,
+            label: args.label.clone(),
+            logprobs: args.logprobs,
+            request_id_prefix: args.get_request_id_prefix(),
+            ready_check_timeout_sec: args.ready_check_timeout_sec,
             extra_headers,
             extra_body,
             metadata,
-            dry_run: cli.dry_run,
+            dry_run: args.dry_run,
             goodput,
             ramp_up,
-            multi_turn: cli.multi_turn,
-            multi_turn_num_turns: cli.multi_turn_num_turns,
+            multi_turn: args.multi_turn,
+            multi_turn_num_turns: args.multi_turn_num_turns,
             multi_turn_min_turns,
             multi_turn_max_turns,
-            sharegpt_multi_turn_max_turns: if cli.multi_turn
-                && cli.dataset_name == DatasetName::ShareGpt
-                && cli.multi_turn_max_turns != 0
+            sharegpt_multi_turn_max_turns: if args.multi_turn
+                && args.dataset_name == DatasetName::ShareGpt
+                && args.multi_turn_max_turns != 0
             {
-                Some(cli.multi_turn_max_turns)
+                Some(args.multi_turn_max_turns)
             } else {
                 None
             },
             per_turn_input_len,
-            multi_turn_concurrency: cli.multi_turn_concurrency,
-            multi_turn_delay_ms: cli.multi_turn_delay_ms,
-            multi_turn_prefix_global_ratio: cli.multi_turn_prefix_global_ratio,
-            multi_turn_prefix_conversation_ratio: cli.multi_turn_prefix_conversation_ratio,
-            speed_bench_config: cli.speed_bench_config,
-            speed_bench_category: cli.speed_bench_category.clone(),
-            speed_bench_max_input_len: cli.speed_bench_max_input_len,
-            hf_split: cli.hf_split.clone(),
-            hf_subset: cli.hf_subset.clone(),
-            hf_output_len: cli.hf_output_len,
-            hf_text_column: cli.hf_text_column.clone(),
-            reset_prefix_cache: cli.reset_prefix_cache,
-            prompt_token_ids: cli.prompt_token_ids,
-            random_mm_base_items_per_request: cli.random_mm_base_items_per_request,
-            random_mm_num_mm_items_range_ratio: cli.random_mm_num_mm_items_range_ratio,
+            multi_turn_concurrency: args.multi_turn_concurrency,
+            multi_turn_delay_ms: args.multi_turn_delay_ms,
+            multi_turn_prefix_global_ratio: args.multi_turn_prefix_global_ratio,
+            multi_turn_prefix_conversation_ratio: args.multi_turn_prefix_conversation_ratio,
+            speed_bench_config: args.speed_bench_config,
+            speed_bench_category: args.speed_bench_category.clone(),
+            speed_bench_max_input_len: args.speed_bench_max_input_len,
+            hf_split: args.hf_split.clone(),
+            hf_subset: args.hf_subset.clone(),
+            hf_output_len: args.hf_output_len,
+            hf_text_column: args.hf_text_column.clone(),
+            reset_prefix_cache: args.reset_prefix_cache,
+            prompt_token_ids: args.prompt_token_ids,
+            random_mm_base_items_per_request: args.random_mm_base_items_per_request,
+            random_mm_num_mm_items_range_ratio: args.random_mm_num_mm_items_range_ratio,
             random_mm_limit,
             random_mm_buckets,
-            enable_multimodal_chat: cli.enable_multimodal_chat,
+            enable_multimodal_chat: args.enable_multimodal_chat,
             lora_modules,
-            lora_assignment: cli.lora_assignment,
+            lora_assignment: args.lora_assignment,
         })
     }
 }
@@ -811,17 +822,17 @@ fn parse_goodput(goodput_args: &Option<Vec<String>>) -> Result<GoodputConfig> {
     Ok(config)
 }
 
-fn parse_ramp_up(cli: &Cli) -> Result<Option<RampUpConfig>> {
-    let strategy = match cli.ramp_up_strategy {
+fn parse_ramp_up(args: &BenchServeArgs) -> Result<Option<RampUpConfig>> {
+    let strategy = match args.ramp_up_strategy {
         None => return Ok(None),
         Some(s) => s,
     };
 
-    let start_rps = cli.ramp_up_start_rps.ok_or_else(|| {
+    let start_rps = args.ramp_up_start_rps.ok_or_else(|| {
         BenchError::Config("--ramp-up-start-rps is required when --ramp-up-strategy is set".into())
     })?;
 
-    let end_rps = cli.ramp_up_end_rps.ok_or_else(|| {
+    let end_rps = args.ramp_up_end_rps.ok_or_else(|| {
         BenchError::Config("--ramp-up-end-rps is required when --ramp-up-strategy is set".into())
     })?;
 
@@ -843,7 +854,21 @@ mod tests {
     use clap::Parser;
 
     use super::*;
-    use crate::cli::Cli;
+    use crate::cli::BenchServeArgs;
+
+    #[derive(Parser)]
+    struct TestCli {
+        #[command(flatten)]
+        args: BenchServeArgs,
+    }
+
+    fn parse_args<I, T>(args: I) -> BenchServeArgs
+    where
+        I: IntoIterator<Item = T>,
+        T: Into<std::ffi::OsString> + Clone,
+    {
+        TestCli::parse_from(args).args
+    }
 
     fn base_multi_turn_args() -> Vec<&'static str> {
         vec![
@@ -859,8 +884,8 @@ mod tests {
     #[test]
     fn test_prefix_sharing_defaults_to_zero() {
         let args = base_multi_turn_args();
-        let cli = Cli::parse_from(args);
-        let config = BenchConfig::from_cli(&cli).unwrap();
+        let args = parse_args(args);
+        let config = BenchConfig::from_args(&args).unwrap();
         assert_eq!(config.multi_turn_prefix_global_ratio, 0.0);
         assert_eq!(config.multi_turn_prefix_conversation_ratio, 0.0);
     }
@@ -874,8 +899,8 @@ mod tests {
             "--multi-turn-prefix-conversation-ratio",
             "0.8",
         ]);
-        let cli = Cli::parse_from(args);
-        let config = BenchConfig::from_cli(&cli).unwrap();
+        let args = parse_args(args);
+        let config = BenchConfig::from_args(&args).unwrap();
         assert!((config.multi_turn_prefix_global_ratio - 0.1).abs() < 1e-10);
         assert!((config.multi_turn_prefix_conversation_ratio - 0.8).abs() < 1e-10);
     }
@@ -889,8 +914,8 @@ mod tests {
             "--multi-turn-prefix-conversation-ratio",
             "0.6",
         ]);
-        let cli = Cli::parse_from(args);
-        assert!(BenchConfig::from_cli(&cli).is_err());
+        let args = parse_args(args);
+        assert!(BenchConfig::from_args(&args).is_err());
     }
 
     #[test]
@@ -902,16 +927,16 @@ mod tests {
             "--multi-turn-prefix-conversation-ratio",
             "0.5",
         ]);
-        let cli = Cli::parse_from(args);
-        assert!(BenchConfig::from_cli(&cli).is_err());
+        let args = parse_args(args);
+        assert!(BenchConfig::from_args(&args).is_err());
     }
 
     #[test]
     fn test_prefix_sharing_out_of_range_fails() {
         let mut args = base_multi_turn_args();
         args.extend(["--multi-turn-prefix-global-ratio", "1.5"]);
-        let cli = Cli::parse_from(args);
-        assert!(BenchConfig::from_cli(&cli).is_err());
+        let args = parse_args(args);
+        assert!(BenchConfig::from_args(&args).is_err());
     }
 
     #[test]
@@ -928,8 +953,8 @@ mod tests {
             "--multi-turn-prefix-global-ratio",
             "0.1",
         ];
-        let cli = Cli::parse_from(args);
-        assert!(BenchConfig::from_cli(&cli).is_err());
+        let args = parse_args(args);
+        assert!(BenchConfig::from_args(&args).is_err());
     }
 
     #[test]
@@ -944,8 +969,8 @@ mod tests {
             "--dataset-name",
             "sharegpt",
         ];
-        let cli = Cli::parse_from(args);
-        let config = BenchConfig::from_cli(&cli).unwrap();
+        let args = parse_args(args);
+        let config = BenchConfig::from_args(&args).unwrap();
 
         assert_eq!(config.multi_turn_max_turns, 3);
         assert_eq!(config.sharegpt_multi_turn_max_turns, None);
@@ -968,8 +993,8 @@ mod tests {
             "--multi-turn-max-turns",
             "2",
         ];
-        let cli = Cli::parse_from(args);
-        let config = BenchConfig::from_cli(&cli).unwrap();
+        let args = parse_args(args);
+        let config = BenchConfig::from_args(&args).unwrap();
         assert_eq!(config.sharegpt_multi_turn_max_turns, Some(2));
     }
 
@@ -987,8 +1012,8 @@ mod tests {
             "--multi-turn-max-turns",
             "1",
         ];
-        let cli = Cli::parse_from(args);
-        let err = BenchConfig::from_cli(&cli).unwrap_err().to_string();
+        let args = parse_args(args);
+        let err = BenchConfig::from_args(&args).unwrap_err().to_string();
         assert!(
             err.contains("at least 2 for ShareGPT"),
             "expected ShareGPT-specific error, got: {err}"
@@ -1009,8 +1034,8 @@ mod tests {
             "--multi-turn-max-turns",
             "20",
         ];
-        let cli = Cli::parse_from(args);
-        let config = BenchConfig::from_cli(&cli).unwrap();
+        let args = parse_args(args);
+        let config = BenchConfig::from_args(&args).unwrap();
 
         assert_eq!(config.sharegpt_multi_turn_max_turns, Some(20));
     }
@@ -1018,8 +1043,8 @@ mod tests {
     #[test]
     fn test_sweep_summary_percentiles_default_empty() {
         let args = base_multi_turn_args();
-        let cli = Cli::parse_from(args);
-        let config = BenchConfig::from_cli(&cli).unwrap();
+        let args = parse_args(args);
+        let config = BenchConfig::from_args(&args).unwrap();
 
         assert!(config.sweep_summary_percentiles.is_empty());
         assert_eq!(config.selected_percentiles, vec![99.0, 90.0]);
@@ -1034,8 +1059,8 @@ mod tests {
             "--sweep-summary-percentiles",
             "90,95,90",
         ]);
-        let cli = Cli::parse_from(args);
-        let config = BenchConfig::from_cli(&cli).unwrap();
+        let args = parse_args(args);
+        let config = BenchConfig::from_args(&args).unwrap();
 
         assert_eq!(config.sweep_summary_percentiles, vec![90.0, 95.0]);
         assert_eq!(config.selected_percentiles, vec![99.0, 95.0, 90.0]);
@@ -1045,8 +1070,8 @@ mod tests {
     fn test_invalid_sweep_summary_percentile_fails() {
         let mut args = base_multi_turn_args();
         args.extend(["--sweep-summary-percentiles", "101"]);
-        let cli = Cli::parse_from(args);
-        assert!(BenchConfig::from_cli(&cli).is_err());
+        let args = parse_args(args);
+        assert!(BenchConfig::from_args(&args).is_err());
     }
 
     #[test]
@@ -1058,12 +1083,20 @@ mod tests {
             "--max-model-len",
             "4096",
         ];
-        let cli = Cli::parse_from(args);
-        let config = BenchConfig::from_cli(&cli).unwrap();
+        let args = parse_args(args);
+        let config = BenchConfig::from_args(&args).unwrap();
 
         assert_eq!(config.max_model_len, Some(4096));
     }
 
+    #[test]
+    fn test_tokenizer_id_deferred_when_model_is_unspecified() {
+        let args = parse_args(["vllm-bench"]);
+        let config = BenchConfig::from_args(&args).unwrap();
+
+        assert_eq!(config.tokenizer_id, None);
+    }
+
     #[test]
     fn test_zero_max_model_len_fails() {
         let args = vec![
@@ -1073,9 +1106,9 @@ mod tests {
             "--max-model-len",
             "0",
         ];
-        let cli = Cli::parse_from(args);
+        let args = parse_args(args);
 
-        assert!(BenchConfig::from_cli(&cli).is_err());
+        assert!(BenchConfig::from_args(&args).is_err());
     }
     #[test]
     fn test_range_ratio_parse_float() {

rust/src/bench/src/datasets/custom.rs

@@ -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]

rust/src/bench/src/datasets/hf_dataset.rs

@@ -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.

rust/src/bench/src/datasets/mod.rs

@@ -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"
     );
 }
 

rust/src/bench/src/datasets/multi_turn.rs

@@ -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

rust/src/bench/src/datasets/prefix_repetition.rs

@@ -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]

rust/src/bench/src/datasets/progress.rs

@@ -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)));
+    }
+}

rust/src/bench/src/datasets/random.rs

@@ -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) => {

rust/src/bench/src/datasets/random_rerank.rs

@@ -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 {

rust/src/bench/src/datasets/sharegpt.rs

@@ -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"
             );
         }
     }

rust/src/bench/src/datasets/speed_bench.rs

@@ -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)
 }

rust/src/bench/src/hub.rs

@@ -1,51 +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 api = hf_hub::api::tokio::Api::new()
-                    .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}"))
     }
 }

rust/src/bench/src/lib.rs

@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+mod backends;
+mod benchmark;
+mod cli;
+mod compare;
+mod config;
+mod datasets;
+mod error;
+mod hub;
+mod metrics;
+mod multi_run;
+mod multi_turn;
+mod output;
+mod rate_control;
+mod ready_checker;
+mod sweep;
+mod tiktoken;
+mod tokenizer;
+
+use anyhow::Context;
+
+pub use cli::{
+    BackendKind, BenchServeArgs, DatasetName, LoraAssignment, RampUpStrategy, SpeedBenchConfig,
+};
+use config::BenchConfig;
+
+/// Prepare process-wide resources for a benchmark run.
+pub fn prepare_process() {
+    // Raise the open-file soft limit to the hard limit. High-concurrency
+    // benchmarks (1024+ requests) easily exceed the default 1024 fd soft limit.
+    if let Ok(new) = rlimit::increase_nofile_limit(u64::MAX)
+        && new > 1024
+    {
+        tracing::info!(soft_limit = new, "raised open-file limit");
+    }
+}
+
+/// Run the online serving benchmark.
+pub async fn run(args: BenchServeArgs) -> anyhow::Result<()> {
+    // --- Compare mode: no server needed, just diff two JSON files ---
+    if let Some(ref files) = args.compare {
+        return compare::compare_results(&files[0], &files[1]).context("Comparison failed");
+    }
+
+    let config = BenchConfig::from_args(&args).context("Configuration error")?;
+
+    async {
+        if config.multi_turn {
+            if let Some(ref sweep_mc) = args.sweep_max_concurrency {
+                // --- Sweep over concurrency in multi-turn mode ---
+                let values = sweep::parse_concurrency_values(sweep_mc)
+                    .context("Invalid --sweep-max-concurrency")?;
+                sweep::run_multi_turn_concurrency_sweep(
+                    &config,
+                    &values,
+                    args.sweep_num_prompts_factor,
+                )
+                .await?;
+            } else {
+                // --- Single multi-turn conversation benchmark ---
+                multi_turn::run_multi_turn_benchmark(&config).await?;
+            }
+        } else if let Some(ref sweep_mc) = args.sweep_max_concurrency {
+            // --- Sweep over max-concurrency ---
+            let values = sweep::parse_concurrency_values(sweep_mc)
+                .context("Invalid --sweep-max-concurrency")?;
+            sweep::run_concurrency_sweep(&config, &values, args.sweep_num_prompts_factor).await?;
+        } else if let Some(ref sweep_rate) = args.sweep_request_rate {
+            // --- Sweep over request-rate ---
+            let values =
+                sweep::parse_rate_values(sweep_rate).context("Invalid --sweep-request-rate")?;
+            sweep::run_rate_sweep(&config, &values).await?;
+        } else if args.num_runs > 1 {
+            // --- Multi-run with statistical aggregation ---
+            multi_run::run_multi(&config, args.num_runs).await?;
+        } else {
+            // --- Normal single benchmark ---
+            benchmark::run_benchmark(&config).await?;
+        }
+        anyhow::Ok(())
+    }
+    .await
+    .context("Benchmark failed")
+}

rust/src/bench/src/main.rs

@@ -1,92 +1,44 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
-mod backends;
-mod benchmark;
-mod cli;
-mod compare;
-mod config;
-mod datasets;
-mod error;
-mod hub;
-mod metrics;
-mod multi_run;
-mod multi_turn;
-mod output;
-mod rate_control;
-mod ready_checker;
-mod sweep;
-mod tiktoken;
-mod tokenizer;
-
 #[cfg(not(target_env = "msvc"))]
 #[global_allocator]
 static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
 
 use anyhow::Context;
 use clap::Parser;
-use cli::Cli;
-use config::BenchConfig;
+
+#[derive(Parser)]
+#[command(
+    name = "vllm-bench",
+    about = "Benchmark online serving throughput",
+    version
+)]
+struct Cli {
+    #[command(flatten)]
+    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<()> {
-    // Raise the open-file soft limit to the hard limit. High-concurrency
-    // benchmarks (1024+ requests) easily exceed the default 1024 fd soft limit.
-    if let Ok(new) = rlimit::increase_nofile_limit(u64::MAX)
-        && new > 1024
-    {
-        eprintln!("Open-file limit: {new}");
-    }
+    init_tracing();
 
     let cli = Cli::parse();
-
-    // --- Compare mode: no server needed, just diff two JSON files ---
-    if let Some(ref files) = cli.compare {
-        return compare::compare_results(&files[0], &files[1]).context("Comparison failed");
-    }
-
-    let config = BenchConfig::from_cli(&cli).context("Configuration error")?;
+    vllm_bench::prepare_process();
 
     let runtime = tokio::runtime::Builder::new_multi_thread()
         .enable_all()
         .build()
-        .expect("Failed to build tokio runtime");
+        .context("Failed to build tokio runtime")?;
 
-    runtime
-        .block_on(async {
-            if config.multi_turn {
-                if let Some(ref sweep_mc) = cli.sweep_max_concurrency {
-                    // --- Sweep over concurrency in multi-turn mode ---
-                    let values = sweep::parse_concurrency_values(sweep_mc)
-                        .context("Invalid --sweep-max-concurrency")?;
-                    sweep::run_multi_turn_concurrency_sweep(
-                        &config,
-                        &values,
-                        cli.sweep_num_prompts_factor,
-                    )
-                    .await?;
-                } else {
-                    // --- Single multi-turn conversation benchmark ---
-                    multi_turn::run_multi_turn_benchmark(&config).await?;
-                }
-            } else if let Some(ref sweep_mc) = cli.sweep_max_concurrency {
-                // --- Sweep over max-concurrency ---
-                let values = sweep::parse_concurrency_values(sweep_mc)
-                    .context("Invalid --sweep-max-concurrency")?;
-                sweep::run_concurrency_sweep(&config, &values, cli.sweep_num_prompts_factor)
-                    .await?;
-            } else if let Some(ref sweep_rate) = cli.sweep_request_rate {
-                // --- Sweep over request-rate ---
-                let values =
-                    sweep::parse_rate_values(sweep_rate).context("Invalid --sweep-request-rate")?;
-                sweep::run_rate_sweep(&config, &values).await?;
-            } else if cli.num_runs > 1 {
-                // --- Multi-run with statistical aggregation ---
-                multi_run::run_multi(&config, cli.num_runs).await?;
-            } else {
-                // --- Normal single benchmark ---
-                benchmark::run_benchmark(&config).await?;
-            }
-            anyhow::Ok(())
-        })
-        .context("Benchmark failed")
+    runtime.block_on(vllm_bench::run(cli.args))
 }

rust/src/bench/src/metrics/calculator.rs

@@ -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> =

rust/src/bench/src/multi_turn.rs

@@ -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");
         }
     }
 

rust/src/bench/src/output/json.rs

@@ -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(())
 }
 

rust/src/bench/src/ready_checker.rs

@@ -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) => {

rust/src/bench/src/sweep.rs

@@ -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();

rust/src/bench/src/tiktoken.rs

@@ -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);
                 }

rust/src/bench/src/tokenizer.rs

@@ -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);
+    }
+}

rust/src/chat/src/multimodal/item.rs

@@ -38,7 +38,7 @@ pub(super) fn build_batched_items(
             let keep_on_cpu = spec.keep_on_cpu_keys.contains(key);
             let (value, field) = match spec.field_layout_for(key) {
                 Some(FieldLayout::Batched) => (
-                    tensor.batched_value_at(index)?,
+                    tensor.batched_wire_value_at(index)?,
                     MmField::Batched(MmBatchedField { keep_on_cpu }),
                 ),
                 Some(FieldLayout::Flat { sizes_key }) => {
@@ -47,7 +47,7 @@ pub(super) fn build_batched_items(
                     })?;
                     let (start, end) = tensor::flat_range_for_index(sizes, sizes_key, index)?;
                     (
-                        tensor.flat_value_range(start, end)?,
+                        tensor.flat_wire_value_range(start, end)?,
                         MmField::Flat(MmFlatField {
                             slices: vec![MmSlice::Slice(SliceSpec {
                                 start: Some(0),
@@ -60,7 +60,7 @@ pub(super) fn build_batched_items(
                     )
                 }
                 None => (
-                    tensor.clone(),
+                    tensor.try_into()?,
                     MmField::Shared(MmSharedField {
                         batch_size: len,
                         keep_on_cpu,
@@ -71,7 +71,7 @@ pub(super) fn build_batched_items(
             data.insert(
                 key.clone(),
                 MmFieldElem {
-                    data: Some(value.try_into()?),
+                    data: Some(value),
                     field,
                 },
             );

rust/src/chat/src/multimodal/tensor.rs

@@ -12,7 +12,7 @@ use vllm_engine_core_client::protocol::tensor::{ShapeExt as _, WireTensor};
 use crate::error::{Error, Result, bail_multimodal, multimodal};
 
 /// Representation for multimodal kwarg values for transformation.
-#[derive(Debug, Clone)]
+#[derive(Debug)]
 pub(super) enum KwargValue {
     /// Float tensor with row-major flat data and shape.
     F32Tensor { data: Vec<f32>, shape: Vec<usize> },
@@ -107,28 +107,19 @@ impl KwargValue {
     }
 }
 
-impl TryFrom<KwargValue> for ProtocolKwargValue {
+impl TryFrom<&KwargValue> for ProtocolKwargValue {
     type Error = Error;
 
-    fn try_from(value: KwargValue) -> Result<Self> {
-        match value {
-            KwargValue::F32Tensor { data, shape } => Ok(Self::Tensor(
-                WireTensor::from_f32(shape, data).map_err(Error::Multimodal)?,
-            )),
-            KwargValue::F16Tensor { data, shape } => Ok(Self::Tensor(
-                WireTensor::from_f16(shape, data).map_err(Error::Multimodal)?,
-            )),
-            KwargValue::Bf16Tensor { data, shape } => Ok(Self::Tensor(
-                WireTensor::from_bf16(shape, data).map_err(Error::Multimodal)?,
-            )),
-            KwargValue::I64Tensor { data, shape } => Ok(Self::Tensor(
-                WireTensor::from_i64(shape, data).map_err(Error::Multimodal)?,
-            )),
-            KwargValue::U32Tensor { data, shape } => Ok(Self::Tensor(
-                WireTensor::from_u32(shape, data).map_err(Error::Multimodal)?,
-            )),
-            KwargValue::Passthrough(value) => Ok(value),
-        }
+    fn try_from(value: &KwargValue) -> Result<Self> {
+        let tensor = match value {
+            KwargValue::F32Tensor { data, shape } => WireTensor::from_f32(shape.clone(), data),
+            KwargValue::F16Tensor { data, shape } => WireTensor::from_f16(shape.clone(), data),
+            KwargValue::Bf16Tensor { data, shape } => WireTensor::from_bf16(shape.clone(), data),
+            KwargValue::I64Tensor { data, shape } => WireTensor::from_i64(shape.clone(), data),
+            KwargValue::U32Tensor { data, shape } => WireTensor::from_u32(shape.clone(), data),
+            KwargValue::Passthrough(value) => return Ok(value.clone()),
+        };
+        tensor.map(ProtocolKwargValue::Tensor).map_err(Error::Multimodal)
     }
 }
 
@@ -145,63 +136,55 @@ impl KwargValue {
         }
     }
 
-    /// Extract one media item from a batched tensor field.
+    /// Convert one media item from a batched tensor field to wire bytes.
     ///
     /// Batched fields use their first axis as media-item index and drop that
     /// axis in the per-feature value, matching vLLM's batched-field semantics.
-    pub(super) fn batched_value_at(&self, index: usize) -> Result<Self> {
-        match self {
-            Self::F32Tensor { data, shape } => {
-                let (shape, data) = slice_first_axis_range(shape, data, index, index + 1, true)?;
-                Ok(Self::F32Tensor { data, shape })
-            }
-            Self::F16Tensor { data, shape } => {
-                let (shape, data) = slice_first_axis_range(shape, data, index, index + 1, true)?;
-                Ok(Self::F16Tensor { data, shape })
-            }
-            Self::Bf16Tensor { data, shape } => {
-                let (shape, data) = slice_first_axis_range(shape, data, index, index + 1, true)?;
-                Ok(Self::Bf16Tensor { data, shape })
-            }
-            Self::I64Tensor { data, shape } => {
-                let (shape, data) = slice_first_axis_range(shape, data, index, index + 1, true)?;
-                Ok(Self::I64Tensor { data, shape })
-            }
-            Self::U32Tensor { data, shape } => {
-                let (shape, data) = slice_first_axis_range(shape, data, index, index + 1, true)?;
-                Ok(Self::U32Tensor { data, shape })
-            }
-            Self::Passthrough(value) => Ok(Self::Passthrough(value.clone())),
-        }
+    pub(super) fn batched_wire_value_at(&self, index: usize) -> Result<ProtocolKwargValue> {
+        self.wire_value_range(index, index + 1, true)
     }
 
-    /// Extract one media item's variable-length range from a flat tensor field.
+    /// Convert one media item's flat tensor range directly to wire bytes.
     ///
     /// Flat fields keep the first axis as the sliced length for this item.
-    pub(super) fn flat_value_range(&self, start: usize, end: usize) -> Result<Self> {
-        match self {
+    pub(super) fn flat_wire_value_range(
+        &self,
+        start: usize,
+        end: usize,
+    ) -> Result<ProtocolKwargValue> {
+        self.wire_value_range(start, end, false)
+    }
+
+    fn wire_value_range(
+        &self,
+        start: usize,
+        end: usize,
+        drop_axis: bool,
+    ) -> Result<ProtocolKwargValue> {
+        let tensor = match self {
             Self::F32Tensor { data, shape } => {
-                let (shape, data) = slice_first_axis_range(shape, data, start, end, false)?;
-                Ok(Self::F32Tensor { data, shape })
+                let (shape, data) = slice_first_axis_range(shape, data, start, end, drop_axis)?;
+                WireTensor::from_f32(shape, data)
             }
             Self::F16Tensor { data, shape } => {
-                let (shape, data) = slice_first_axis_range(shape, data, start, end, false)?;
-                Ok(Self::F16Tensor { data, shape })
+                let (shape, data) = slice_first_axis_range(shape, data, start, end, drop_axis)?;
+                WireTensor::from_f16(shape, data)
             }
             Self::Bf16Tensor { data, shape } => {
-                let (shape, data) = slice_first_axis_range(shape, data, start, end, false)?;
-                Ok(Self::Bf16Tensor { data, shape })
+                let (shape, data) = slice_first_axis_range(shape, data, start, end, drop_axis)?;
+                WireTensor::from_bf16(shape, data)
             }
             Self::I64Tensor { data, shape } => {
-                let (shape, data) = slice_first_axis_range(shape, data, start, end, false)?;
-                Ok(Self::I64Tensor { data, shape })
+                let (shape, data) = slice_first_axis_range(shape, data, start, end, drop_axis)?;
+                WireTensor::from_i64(shape, data)
             }
             Self::U32Tensor { data, shape } => {
-                let (shape, data) = slice_first_axis_range(shape, data, start, end, false)?;
-                Ok(Self::U32Tensor { data, shape })
+                let (shape, data) = slice_first_axis_range(shape, data, start, end, drop_axis)?;
+                WireTensor::from_u32(shape, data)
             }
-            Self::Passthrough(value) => Ok(Self::Passthrough(value.clone())),
-        }
+            Self::Passthrough(value) => return Ok(value.clone()),
+        };
+        tensor.map(ProtocolKwargValue::Tensor).map_err(Error::Multimodal)
     }
 }
 
@@ -240,13 +223,13 @@ fn tensor_as_usize_vec(tensor: &KwargValue) -> Result<Vec<usize>> {
 }
 
 /// Slice a flat row-major tensor along its first axis.
-fn slice_first_axis_range<T: Clone>(
+fn slice_first_axis_range<'a, T>(
     shape: &[usize],
-    data: &[T],
+    data: &'a [T],
     start: usize,
     end: usize,
     drop_axis: bool,
-) -> Result<(Vec<usize>, Vec<T>)> {
+) -> Result<(Vec<usize>, &'a [T])> {
     let first_dim = *shape.first().ok_or_else(|| multimodal!("tensor has no first dimension"))?;
     if start > end || end > first_dim {
         bail_multimodal!("invalid tensor slice {start}..{end} for first dimension {first_dim}");
@@ -270,7 +253,7 @@ fn slice_first_axis_range<T: Clone>(
         shape[0] = end - start;
         shape
     };
-    Ok((out_shape, data[data_start..data_end].to_vec()))
+    Ok((out_shape, &data[data_start..data_end]))
 }
 
 #[cfg(test)]
@@ -278,35 +261,39 @@ mod tests {
     use super::*;
 
     #[test]
-    fn batched_value_at_drops_first_axis() {
+    fn batched_wire_value_at_drops_first_axis() {
         let value = KwargValue::F32Tensor {
             data: vec![1.0, 2.0, 3.0, 4.0],
             shape: vec![2, 2],
         };
 
-        let value = value.batched_value_at(1).unwrap();
+        let ProtocolKwargValue::Tensor(tensor) = value.batched_wire_value_at(1).unwrap() else {
+            panic!("expected tensor");
+        };
 
-        assert!(matches!(
-            value,
-            KwargValue::F32Tensor { data, shape }
-                if shape == vec![2] && data == vec![3.0, 4.0]
-        ));
+        assert_eq!(tensor.shape, vec![2]);
+        assert_eq!(
+            tensor.data.into_raw_view().unwrap(),
+            [3.0_f32, 4.0].into_iter().flat_map(f32::to_ne_bytes).collect::<Vec<_>>()
+        );
     }
 
     #[test]
-    fn flat_value_range_keeps_first_axis() {
+    fn flat_wire_value_range_keeps_first_axis() {
         let value = KwargValue::U32Tensor {
             data: (0..10).collect(),
             shape: vec![5, 2],
         };
 
-        let value = value.flat_value_range(1, 3).unwrap();
+        let ProtocolKwargValue::Tensor(tensor) = value.flat_wire_value_range(1, 3).unwrap() else {
+            panic!("expected tensor");
+        };
 
-        assert!(matches!(
-            value,
-            KwargValue::U32Tensor { data, shape }
-                if shape == vec![2, 2] && data == vec![2, 3, 4, 5]
-        ));
+        assert_eq!(tensor.shape, vec![2, 2]);
+        assert_eq!(
+            tensor.data.into_raw_view().unwrap(),
+            [2_u32, 3, 4, 5].into_iter().flat_map(u32::to_ne_bytes).collect::<Vec<_>>()
+        );
     }
 
     #[test]
@@ -336,7 +323,7 @@ mod tests {
         let value =
             KwargValue::from_f32_tensor(vec![1.0, -1.0], vec![2], ModelDtype::BFloat16).unwrap();
 
-        let ProtocolKwargValue::Tensor(tensor) = ProtocolKwargValue::try_from(value).unwrap()
+        let ProtocolKwargValue::Tensor(tensor) = ProtocolKwargValue::try_from(&value).unwrap()
         else {
             panic!("expected tensor");
         };
@@ -351,7 +338,7 @@ mod tests {
         let value =
             KwargValue::from_f32_tensor(vec![1.0, -1.0], vec![2], ModelDtype::Float16).unwrap();
 
-        let ProtocolKwargValue::Tensor(tensor) = ProtocolKwargValue::try_from(value).unwrap()
+        let ProtocolKwargValue::Tensor(tensor) = ProtocolKwargValue::try_from(&value).unwrap()
         else {
             panic!("expected tensor");
         };

rust/src/chat/src/multimodal/video.rs

@@ -130,7 +130,7 @@ fn build_video_item(
         let keep_on_cpu = support.spec.keep_on_cpu_keys.contains(&key);
         let (value, field) = match support.spec.field_layout_for(&key) {
             Some(FieldLayout::Batched) => (
-                tensor.batched_value_at(0)?,
+                tensor.batched_wire_value_at(0)?,
                 MmField::Batched(MmBatchedField { keep_on_cpu }),
             ),
             Some(FieldLayout::Flat { .. }) => {
@@ -138,7 +138,7 @@ fn build_video_item(
                     .first_dim()
                     .ok_or_else(|| multimodal!("flat video input `{key}` is not a tensor"))?;
                 (
-                    tensor,
+                    (&tensor).try_into()?,
                     MmField::Flat(MmFlatField {
                         slices: vec![MmSlice::Slice(SliceSpec {
                             start: Some(0),
@@ -151,7 +151,7 @@ fn build_video_item(
                 )
             }
             None => (
-                tensor,
+                (&tensor).try_into()?,
                 MmField::Shared(MmSharedField {
                     batch_size: 1,
                     keep_on_cpu,
@@ -162,7 +162,7 @@ fn build_video_item(
         data.insert(
             key,
             MmFieldElem {
-                data: Some(value.try_into()?),
+                data: Some(value),
                 field,
             },
         );

rust/src/chat/src/output/default/mod.rs

@@ -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;

rust/src/chat/src/output/default/structural_tag.rs

@@ -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);

rust/src/chat/src/renderer/hf/format.rs

@@ -236,9 +236,10 @@ fn has_content_item_loop(root: &Stmt<'_>) -> bool {
 
     loops.into_iter().any(|loop_ast| {
         matches!(loop_ast.target, Expr::Var(_))
-            && message_varnames
-                .iter()
-                .any(|varname| is_var_or_elems_access(&loop_ast.iter, varname, Some("content")))
+            && (is_var_access(&loop_ast.iter, "content")
+                || message_varnames.iter().any(|varname| {
+                    is_var_or_elems_access(&loop_ast.iter, varname, Some("content"))
+                }))
     })
 }
 
@@ -315,6 +316,16 @@ mod tests {
         );
     }
 
+    #[test]
+    fn detects_openai_template_with_content_parameter_loop() {
+        assert_eq!(
+            detect(
+                "{% macro render(content) %}{% for item in content %}{{ item }}{% endfor %}{% endmacro %}{% for message in messages %}{{ render(message.content) }}{% endfor %}"
+            ),
+            ChatTemplateContentFormat::OpenAi
+        );
+    }
+
     #[test]
     fn detects_openai_template_with_messages_alias() {
         assert_eq!(

rust/src/chat/src/renderer/hf/mod.rs

@@ -1309,6 +1309,26 @@ mod tests {
         .assert_eq(&rendered);
     }
 
+    #[test]
+    fn qwen35_template_auto_detects_openai_multimodal_content() {
+        let mut request = image_request();
+        request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
+
+        let rendered = render_mm(
+            QWEN3_5_0_8B_TEMPLATE,
+            &request,
+            ChatTemplateContentFormatOption::Auto,
+        )
+        .unwrap();
+
+        expect![[r#"
+            Text(
+                "<|im_start|>user\na<|vision_start|><|image_pad|><|vision_end|>b<|im_end|>\n",
+            )
+        "#]]
+        .assert_debug_eq(&rendered.prompt);
+    }
+
     #[test]
     fn qwen35_template_renders_closed_empty_reasoning_span_when_thinking_disabled() {
         let mut request = sample_request(vec![ChatMessage::text(ChatRole::User, "hello")]);

rust/src/cmd/Cargo.toml

@@ -29,6 +29,7 @@ tokio-util.workspace = true
 tracing.workspace = true
 tracing-subscriber.workspace = true
 uuid.workspace = true
+vllm-bench.workspace = true
 vllm-chat.workspace = true
 vllm-engine-core-client.workspace = true
 vllm-managed-engine.workspace = true

rust/src/cmd/src/cli.rs

@@ -79,13 +79,23 @@ impl Cli {
 }
 
 /// Supported top-level CLI commands.
-#[derive(Debug, Subcommand, PartialEq, Eq)]
+#[derive(Debug, Subcommand)]
 pub enum Command {
     /// Run the Rust OpenAI frontend as a Python-supervised worker.
     Frontend(FrontendArgs),
     /// Launch a managed Python headless engine, then run the Rust OpenAI
     /// frontend.
     Serve(ServeArgs),
+    /// Run vLLM benchmarks.
+    #[command(subcommand)]
+    Bench(BenchCommand),
+}
+
+/// Supported benchmark commands.
+#[derive(Debug, Subcommand)]
+pub enum BenchCommand {
+    /// Benchmark online serving throughput.
+    Serve(vllm_bench::BenchServeArgs),
 }
 
 /// A JSON-encoded list of strings, matching Python's `json.loads` CLI type for

rust/src/cmd/src/cli/tests.rs

@@ -5,7 +5,27 @@ use expect_test::expect;
 use vllm_engine_core_client::TransportMode;
 use vllm_server::{Config, HttpListenerMode, ParserSelection, RendererSelection};
 
-use super::{Cli, Command};
+use super::{BenchCommand, Cli, Command};
+
+#[test]
+fn bench_serve_args_parse_without_managed_engine_repartition() {
+    let cli = Cli::try_parse_from([
+        "vllm-rs",
+        "bench",
+        "serve",
+        "--backend",
+        "openai-chat",
+        "--request-rate",
+        "inf",
+    ])
+    .unwrap();
+
+    let Command::Bench(BenchCommand::Serve(args)) = cli.command else {
+        panic!("expected bench serve args");
+    };
+    assert_eq!(args.backend, vllm_bench::BackendKind::OpenaiChat);
+    assert!(args.request_rate.is_infinite());
+}
 
 #[test]
 fn serve_args_forward_python_flags_with_separator() {

rust/src/cmd/src/logging.rs

@@ -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"));

rust/src/cmd/src/main.rs

@@ -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};
@@ -12,7 +13,7 @@ use tokio_util::sync::CancellationToken;
 use tracing::{info, warn};
 use vllm_managed_engine::ManagedEngineHandle;
 
-use crate::cli::{Cli, Command};
+use crate::cli::{BenchCommand, Cli, Command};
 
 #[global_allocator]
 static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
@@ -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();
@@ -100,6 +108,10 @@ fn main() -> Result<()> {
 async fn async_main(cli: Cli) -> Result<()> {
     match cli.command {
         Command::Frontend(args) => vllm_server::serve(args.into_config(), shutdown_signal()).await,
+        Command::Bench(BenchCommand::Serve(bench_args)) => {
+            vllm_bench::prepare_process();
+            vllm_bench::run(bench_args).await
+        }
         Command::Serve(args) => {
             let handshake_port = args.managed_engine.resolve_handshake_port()?;
 

rust/src/engine-core-client/src/client.rs

@@ -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()

rust/src/engine-core-client/src/client/imp.rs

@@ -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!(

rust/src/engine-core-client/src/protocol/logprobs.rs

@@ -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

rust/src/engine-core-client/src/protocol/logprobs/tests.rs

@@ -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"
+    );
+}

rust/src/engine-core-client/src/protocol/tensor.rs

@@ -55,52 +55,57 @@ pub struct WireNdArray {
 
 impl WireNdArray {
     /// Build a float32 tensor/ndarray backed by native-endian raw-view bytes.
-    pub fn from_f32(shape: Vec<usize>, data: Vec<f32>) -> Result<Self, String> {
+    pub fn from_f32(shape: Vec<usize>, data: impl AsRef<[f32]>) -> Result<Self, String> {
+        let data = data.as_ref();
         validate_element_count(&shape, data.len())?;
         Ok(Self {
             dtype: "float32".to_string(),
             shape,
-            data: WireArrayData::RawView(pod_collect_to_vec::<f32, u8>(&data)),
+            data: WireArrayData::RawView(pod_collect_to_vec::<f32, u8>(data)),
         })
     }
 
     /// Build a float16 tensor/ndarray backed by native-endian raw-view bytes.
-    pub fn from_f16(shape: Vec<usize>, data: Vec<f16>) -> Result<Self, String> {
+    pub fn from_f16(shape: Vec<usize>, data: impl AsRef<[f16]>) -> Result<Self, String> {
+        let data = data.as_ref();
         validate_element_count(&shape, data.len())?;
         Ok(Self {
             dtype: "float16".to_string(),
             shape,
-            data: WireArrayData::RawView(pod_collect_to_vec::<f16, u8>(&data)),
+            data: WireArrayData::RawView(pod_collect_to_vec::<f16, u8>(data)),
         })
     }
 
     /// Build a bfloat16 tensor/ndarray backed by native-endian raw-view bytes.
-    pub fn from_bf16(shape: Vec<usize>, data: Vec<bf16>) -> Result<Self, String> {
+    pub fn from_bf16(shape: Vec<usize>, data: impl AsRef<[bf16]>) -> Result<Self, String> {
+        let data = data.as_ref();
         validate_element_count(&shape, data.len())?;
         Ok(Self {
             dtype: "bfloat16".to_string(),
             shape,
-            data: WireArrayData::RawView(pod_collect_to_vec::<bf16, u8>(&data)),
+            data: WireArrayData::RawView(pod_collect_to_vec::<bf16, u8>(data)),
         })
     }
 
     /// Build an int64 tensor/ndarray backed by native-endian raw-view bytes.
-    pub fn from_i64(shape: Vec<usize>, data: Vec<i64>) -> Result<Self, String> {
+    pub fn from_i64(shape: Vec<usize>, data: impl AsRef<[i64]>) -> Result<Self, String> {
+        let data = data.as_ref();
         validate_element_count(&shape, data.len())?;
         Ok(Self {
             dtype: "int64".to_string(),
             shape,
-            data: WireArrayData::RawView(pod_collect_to_vec::<i64, u8>(&data)),
+            data: WireArrayData::RawView(pod_collect_to_vec::<i64, u8>(data)),
         })
     }
 
     /// Build a uint32 tensor/ndarray backed by native-endian raw-view bytes.
-    pub fn from_u32(shape: Vec<usize>, data: Vec<u32>) -> Result<Self, String> {
+    pub fn from_u32(shape: Vec<usize>, data: impl AsRef<[u32]>) -> Result<Self, String> {
+        let data = data.as_ref();
         validate_element_count(&shape, data.len())?;
         Ok(Self {
             dtype: "uint32".to_string(),
             shape,
-            data: WireArrayData::RawView(pod_collect_to_vec::<u32, u8>(&data)),
+            data: WireArrayData::RawView(pod_collect_to_vec::<u32, u8>(data)),
         })
     }
 

rust/src/parser/benches/utils/adapter.rs

@@ -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> {

rust/src/parser/src/tool/deepseek_dsml/deepseek_v32.rs

@@ -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<()> {

rust/src/parser/src/tool/deepseek_dsml/deepseek_v4.rs

@@ -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]

rust/src/parser/src/tool/deepseek_json/deepseek_v3.rs

@@ -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<()> {

rust/src/parser/src/tool/deepseek_json/deepseek_v31.rs

@@ -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<()> {

rust/src/parser/src/tool/glm_xml/glm47_moe.rs

@@ -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<()> {

rust/src/parser/src/tool/hy_v3.rs

@@ -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<()> {

rust/src/parser/src/tool/json/hermes.rs

@@ -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<()> {

rust/src/parser/src/tool/json/llama.rs

@@ -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<()> {

rust/src/parser/src/tool/json/qwen.rs

@@ -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<()> {

rust/src/parser/src/tool/kimi_k2.rs

@@ -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> {

rust/src/parser/src/tool/minimax_m2.rs

@@ -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<()> {

rust/src/parser/src/tool/mod.rs

@@ -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
     }
 

rust/src/parser/src/tool/qwen_coder.rs

@@ -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]

rust/src/parser/src/unified/combined.rs

@@ -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,

rust/src/parser/src/unified/mod.rs

@@ -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
     }
 

rust/src/server/Cargo.toml

@@ -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

rust/src/server/src/grpc/health.rs

@@ -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;
+}

rust/src/server/src/grpc/mod.rs

@@ -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;
 

rust/src/server/src/grpc/tests.rs

@@ -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");
+}

rust/src/server/src/lib.rs

@@ -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
         }
     };

rust/src/server/src/routes/inference/generate.rs

@@ -238,13 +238,18 @@ fn collect_generate(
         None
     };
     let prompt_logprobs = if include_prompt_logprobs {
-        let prompt_logprobs = collected.prompt_logprobs.as_ref().ok_or_else(|| {
-            ApiError::server_error(
-                "raw generate response requested prompt_logprobs but generation returned none"
-                    .to_string(),
-            )
-        })?;
-        Some(raw_prompt_logprobs_to_maps(prompt_logprobs))
+        match collected.prompt_logprobs.as_ref() {
+            Some(prompt_logprobs) => Some(raw_prompt_logprobs_to_maps(prompt_logprobs)),
+            // A single-token prompt has no scored positions; same mapping
+            // as /v1/completions.
+            None if collected.prompt_token_ids.len() == 1 => Some(vec![None]),
+            None => {
+                return Err(ApiError::server_error(
+                    "raw generate response requested prompt_logprobs but generation returned none"
+                        .to_string(),
+                ));
+            }
+        }
     } else {
         None
     };
@@ -472,4 +477,48 @@ mod tests {
             Some(2)
         );
     }
+
+    #[test]
+    fn collect_generate_maps_prompt_logprobs_for_single_token_prompt() {
+        let output_without_payload = |prompt_token_ids: Vec<u32>| CollectedGenerateOutput {
+            request_id: "raw-1".to_string(),
+            prompt_logprobs: None,
+            token_ids: vec![3],
+            logprobs: None,
+            finish_reason: FinishReason::stop_eos(),
+            usage: vllm_llm::TokenUsage {
+                prompt_token_count: prompt_token_ids.len(),
+                output_token_count: 1,
+                cached_token_count: 0,
+            },
+            kv_transfer_params: None,
+            ec_transfer_params: None,
+            prompt_token_ids,
+        };
+
+        let response = collect_generate(
+            output_without_payload(vec![9707]),
+            "raw-1".to_string(),
+            ApiServerOptions::default(),
+            ResponseOptions {
+                include_prompt_logprobs: true,
+                ..Default::default()
+            },
+        )
+        .expect("single-token prompt without payload maps to [None]");
+        let prompt_logprobs = response.prompt_logprobs.expect("prompt logprobs present");
+        assert_eq!(prompt_logprobs.len(), 1);
+        assert!(prompt_logprobs[0].is_none());
+
+        collect_generate(
+            output_without_payload(vec![9707, 11]),
+            "raw-2".to_string(),
+            ApiServerOptions::default(),
+            ResponseOptions {
+                include_prompt_logprobs: true,
+                ..Default::default()
+            },
+        )
+        .expect_err("multi-token prompt without payload is an engine failure");
+    }
 }

rust/src/server/src/routes/openai/chat_completions.rs

@@ -35,7 +35,7 @@ use crate::routes::openai::chat_completions::types::{
     ChatMessageDelta,
 };
 use crate::routes::openai::utils::logprobs::{
-    decoded_logprobs_to_openai_chat, decoded_prompt_logprobs_to_maps,
+    decoded_logprobs_to_openai_chat, prompt_logprobs_to_maps,
 };
 use crate::routes::openai::utils::types::{
     ChatLogProbs, FunctionCallDelta, FunctionCallResponse, ToolCall, ToolCallDelta, Usage,
@@ -181,14 +181,11 @@ async fn collect_chat_completion(
         None
     };
     let prompt_logprobs = if include_prompt_logprobs {
-        Some(decoded_prompt_logprobs_to_maps(
-            prompt_logprobs.as_ref().ok_or_else(|| {
-                server_error!(
-                    "chat response requested prompt_logprobs but generation returned none"
-                )
-            })?,
+        Some(prompt_logprobs_to_maps(
+            prompt_logprobs.as_ref(),
+            &prompt_token_ids,
             return_tokens_as_token_ids,
-        ))
+        )?)
     } else {
         None
     };

rust/src/server/src/routes/openai/completions.rs

@@ -5,7 +5,6 @@ mod convert;
 mod types;
 mod validate;
 
-use std::collections::HashMap;
 use std::convert::Infallible;
 use std::result::Result;
 use std::sync::Arc;
@@ -29,8 +28,8 @@ use vllm_text::{
 
 use self::convert::{ResponseOptions, prepare_completion_request};
 use super::utils::logprobs::{
-    collected_logprobs_to_openai, decoded_logprobs_to_openai, decoded_prompt_logprobs_to_maps,
-    decoded_prompt_logprobs_to_openai, text_len,
+    collected_logprobs_to_openai, decoded_logprobs_to_openai, decoded_prompt_logprobs_to_openai,
+    prompt_logprobs_to_maps, text_len,
 };
 use super::utils::types::Usage;
 use crate::config::ApiServerOptions;
@@ -505,27 +504,6 @@ fn prompt_only_logprobs_to_openai(
     ))
 }
 
-fn prompt_logprobs_to_maps(
-    prompt_logprobs: Option<&DecodedPromptLogprobs>,
-    prompt_token_ids: &[u32],
-    return_tokens_as_token_ids: bool,
-) -> Result<Vec<Option<HashMap<String, f32>>>, ApiError> {
-    if let Some(prompt_logprobs) = prompt_logprobs {
-        return Ok(decoded_prompt_logprobs_to_maps(
-            prompt_logprobs,
-            return_tokens_as_token_ids,
-        ));
-    }
-
-    if let [_token_id] = prompt_token_ids {
-        return Ok(vec![None]);
-    }
-
-    Err(server_error!(
-        "completion response requested prompt_logprobs but generation returned none"
-    ))
-}
-
 fn usage_chunk(
     request_id: &str,
     response_model: &str,

rust/src/server/src/routes/openai/utils/logprobs.rs

@@ -100,20 +100,31 @@ pub fn decoded_prompt_logprobs_to_openai(
     })
 }
 
-/// Convert decoded prompt logprobs into the vLLM-style prompt-logprobs response
-/// shape.
-pub fn decoded_prompt_logprobs_to_maps(
-    prompt_logprobs: &DecodedPromptLogprobs,
+/// Map decoded prompt logprobs into vLLM-style per-position maps, treating a
+/// missing single-token payload as `[None]`.
+pub fn prompt_logprobs_to_maps(
+    prompt_logprobs: Option<&DecodedPromptLogprobs>,
+    prompt_token_ids: &[u32],
     return_tokens_as_token_ids: bool,
-) -> Vec<Option<HashMap<String, f32>>> {
-    std::iter::once(None)
-        .chain(prompt_logprobs.scored_positions.iter().map(|position| {
-            Some(position_top_logprobs_map(
-                position,
-                return_tokens_as_token_ids,
-            ))
-        }))
-        .collect()
+) -> Result<Vec<Option<HashMap<String, f32>>>, ApiError> {
+    if let Some(prompt_logprobs) = prompt_logprobs {
+        return Ok(std::iter::once(None)
+            .chain(prompt_logprobs.scored_positions.iter().map(|position| {
+                Some(position_top_logprobs_map(
+                    position,
+                    return_tokens_as_token_ids,
+                ))
+            }))
+            .collect());
+    }
+
+    if let [_token_id] = prompt_token_ids {
+        return Ok(vec![None]);
+    }
+
+    Err(server_error!(
+        "prompt_logprobs were requested but generation returned none"
+    ))
 }
 
 /// Convert decoded token-position logprobs into the OpenAI chat `logprobs`
@@ -275,7 +286,13 @@ pub fn clamp_logprob(logprob: f32) -> f32 {
 mod tests {
     use vllm_text::{DecodedLogprobs, DecodedPositionLogprobs, DecodedTokenLogprob};
 
-    use super::decoded_logprobs_to_openai_chat;
+    use super::{decoded_logprobs_to_openai_chat, prompt_logprobs_to_maps};
+
+    #[test]
+    fn prompt_logprobs_maps_reject_missing_multi_token_payload() {
+        prompt_logprobs_to_maps(None, &[9707, 11], false)
+            .expect_err("multi-token prompt without payload is an engine failure");
+    }
 
     fn sample_logprobs() -> DecodedLogprobs {
         DecodedLogprobs {

tests/compile/passes/test_fusion_attn.py

@@ -39,7 +39,12 @@ from vllm.platforms import current_platform
 from vllm.utils.flashinfer import has_flashinfer
 from vllm.v1.attention.backend import AttentionMetadata
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
-from vllm.v1.kv_cache_interface import AttentionSpec, get_kv_quant_mode
+from vllm.v1.attention.backends.utils import resolve_kv_cache_layout
+from vllm.v1.kv_cache_interface import (
+    AttentionSpec,
+    get_kv_quant_mode,
+    reshape_kv_cache,
+)
 
 DEVICE_TYPE = current_platform.device_type
 FP8_DTYPE = current_platform.fp8_dtype()
@@ -108,32 +113,27 @@ class AttentionQuantPatternModel(torch.nn.Module):
         max_blocks = (max(batch_spec.seq_lens) + self.block_size - 1) // self.block_size
         num_blocks = batch_size * max_blocks
 
-        # Fetch the attention backend and kv cache shape and stride order
-        attn_backend = self.attn.attn_backend
-        kv_cache_shape = attn_backend.get_kv_cache_shape(
-            num_blocks,
-            self.block_size,
-            self.num_kv_heads,
-            self.head_size,
-            cache_dtype_str=self.attn.kv_cache_dtype,
-        )
-        try:
-            kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
-        except (AttributeError, NotImplementedError):
-            kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
-
-        kv_cache_shape = tuple(kv_cache_shape[i] for i in kv_cache_stride_order)
-        inv_order = [
-            kv_cache_stride_order.index(i) for i in range(len(kv_cache_stride_order))
-        ]
-
-        # Create dummy KV cache
-        raw_tensor = torch.zeros(
-            kv_cache_shape,
+        spec = AttentionSpec(
+            block_size=self.block_size,
+            num_kv_heads=self.num_kv_heads,
+            head_size=self.head_size,
             dtype=self.attn.kv_cache_torch_dtype,
+            kv_quant_mode=get_kv_quant_mode(self.attn.kv_cache_dtype),
+        )
+        layout = resolve_kv_cache_layout()
+        num_layer_slots = 1 if layout.is_layer_compact else 2
+        raw_tensor = torch.zeros(
+            num_layer_slots * num_blocks * spec.page_size_bytes,
+            dtype=torch.int8,
             device=self.device,
         )
-        kv_cache = raw_tensor.permute(*inv_order)
+        kv_cache = reshape_kv_cache(
+            raw_tensor,
+            spec,
+            num_blocks,
+            num_layer_slots,
+            layout,
+        )[0]
 
         self.attn.kv_cache = kv_cache
 

tests/compile/passes/test_mla_attn_quant_fusion.py

@@ -150,27 +150,14 @@ class MLAAttentionQuantPatternModel(torch.nn.Module):
         max_blocks = (max(batch_spec.seq_lens) + self.block_size - 1) // self.block_size
         num_blocks = batch_size * max_blocks
 
-        # MLA KV cache is 3D: (num_blocks, block_size, head_size)
-        attn_backend = self.mla_attn.attn_backend
-        kv_cache_shape = attn_backend.get_kv_cache_shape(
-            num_blocks, self.block_size, 1, self.head_size
+        # MLA KV cache is 4D: (num_blocks, num_heads=1, block_size, head_size)
+        kv_cache = torch.zeros(
+            (num_blocks, 1, self.block_size, self.head_size),
+            dtype=self.kv_cache_dtype,
+            device=self.device,
         )
-        try:
-            kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
-        except (AttributeError, NotImplementedError):
-            kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
 
-        ordered_shape = tuple(kv_cache_shape[i] for i in kv_cache_stride_order)
-        inv_order = [
-            kv_cache_stride_order.index(i) for i in range(len(kv_cache_stride_order))
-        ]
-
-        raw_tensor = torch.zeros(
-            ordered_shape, dtype=self.kv_cache_dtype, device=self.device
-        )
-        kv_cache = raw_tensor.permute(*inv_order)
-
-        self.mla_attn.kv_cache = kv_cache
+        self.mla_attn.bind_kv_cache(kv_cache)
 
         self.attn_metadata = self.builder.build(
             common_prefix_len=0, common_attn_metadata=common_attn_metadata

tests/compile/passes/test_mla_rope_kvcache_cat_fusion.py

@@ -165,29 +165,15 @@ class MLARoPEKVCacheCatTestModel(torch.nn.Module):
         max_blocks = (max(batch_spec.seq_lens) + self.block_size - 1) // self.block_size
         num_blocks = batch_size * max_blocks
 
-        # Fetch the attention backend and kv cache shape and stride order
-        kv_cache_shape = self.attn_backend.get_kv_cache_shape(
-            num_blocks, self.block_size, self.num_kv_heads, self.head_size
-        )
-        try:
-            kv_cache_stride_order = self.attn_backend.get_kv_cache_stride_order()
-        except (AttributeError, NotImplementedError):
-            kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
-
-        kv_cache_shape = tuple(kv_cache_shape[i] for i in kv_cache_stride_order)
-        inv_order = [
-            kv_cache_stride_order.index(i) for i in range(len(kv_cache_stride_order))
-        ]
-
-        raw_tensor = torch.zeros(
-            num_blocks * self.block_size * self.num_kv_heads * self.head_size,
+        # MLA uses a 4D KV cache: (num_blocks, num_heads=1, block_size, head_size).
+        kv_cache_shape = (num_blocks, 1, self.block_size, self.head_size)
+        kv_cache = torch.zeros(
+            kv_cache_shape,
             dtype=self.kv_cache_dtype,
             device=self.device,
         )
-        raw_tensor = raw_tensor.view(kv_cache_shape)
-        kv_cache = raw_tensor.permute(*inv_order)
 
-        self.mla_attn.kv_cache = kv_cache
+        self.mla_attn.bind_kv_cache(kv_cache)
 
         # Build attn metadata
         attn_metadata = self.builder.build(

tests/compile/passes/test_rocm_aiter_qk_norm_rope_kvcache_fusion.py

@@ -38,7 +38,7 @@ from vllm.v1.attention.backend import (
     CommonAttentionMetadata,
 )
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
-from vllm.v1.kv_cache_interface import AttentionSpec
+from vllm.v1.kv_cache_interface import AttentionSpec, KVCacheLayout, reshape_kv_cache
 
 INDEX_SELECT_OP = torch.ops.aten.index.Tensor
 FP8_DTYPE = current_platform.fp8_dtype()
@@ -128,20 +128,21 @@ class QKNormRoPEKVCacheTestModel(torch.nn.Module):
             self.attn._k_scale = self.attn._k_scale.to(device)
             self.attn._v_scale = self.attn._v_scale.to(device)
 
+        self.kv_cache_spec = AttentionSpec(
+            block_size=self.block_size,
+            num_kv_heads=self.num_kv_heads,
+            head_size=head_size,
+            dtype=self.kv_cache_dtype,
+        )
         self.builder = self.attn.attn_backend.get_builder_cls()(
-            kv_cache_spec=AttentionSpec(
-                block_size=self.block_size,
-                num_kv_heads=self.num_kv_heads,
-                head_size=head_size,
-                dtype=self.kv_cache_dtype,
-            ),
+            kv_cache_spec=self.kv_cache_spec,
             layer_names=[self.attn.layer_name],
             vllm_config=vllm_config,
             device=device,
         )
 
     def build_attn_metadata(
-        self, batch_size: int, kv_stride_order: tuple[int, ...] | None = None
+        self, batch_size: int, layout: KVCacheLayout
     ) -> CommonAttentionMetadata:
         batch_spec = BatchSpec(seq_lens=[1] * batch_size, query_lens=[1] * batch_size)
         common_attn_metadata = create_common_attn_metadata(
@@ -151,32 +152,22 @@ class QKNormRoPEKVCacheTestModel(torch.nn.Module):
         max_blocks = (max(batch_spec.seq_lens) + self.block_size - 1) // self.block_size
         num_blocks = batch_size * max_blocks
 
-        attn_backend = self.attn.attn_backend
-        kv_cache_shape = attn_backend.get_kv_cache_shape(
-            num_blocks, self.block_size, self.num_kv_heads, self.head_size
-        )
-        # Caller can force a physical layout; else use the backend's.
-        if kv_stride_order is None:
-            try:
-                kv_stride_order = attn_backend.get_kv_cache_stride_order()
-            except (AttributeError, NotImplementedError):
-                kv_stride_order = tuple(range(len(kv_cache_shape)))
-
-        kv_cache_shape = tuple(kv_cache_shape[i] for i in kv_stride_order)
-        inv_order = [kv_stride_order.index(i) for i in range(len(kv_stride_order))]
-
         raw_tensor = torch.zeros(
-            2 * num_blocks * self.block_size * self.num_kv_heads * self.head_size,
-            dtype=self.kv_cache_dtype,
+            num_blocks * self.kv_cache_spec.page_size_bytes,
+            dtype=torch.int8,
             device=self.device,
         )
-        raw_tensor = raw_tensor.view(kv_cache_shape)
-        kv_cache = raw_tensor.permute(*inv_order)
+        kv_cache = reshape_kv_cache(
+            raw_tensor,
+            self.kv_cache_spec,
+            num_blocks,
+            num_layer_slots=1,
+            layout=layout,
+        )[0]
 
         # Store as a bare tensor (not wrapped in a list) to match production
-        # `bind_kv_cache` behavior.  `get_attention_context` returns this
-        # attribute directly to the fused/unfused `do_kv_cache_update` impls,
-        # which call `kv_cache.unbind(0)` and therefore require a tensor.
+        # `bind_kv_cache` behavior. `get_attention_context` returns this
+        # attribute directly to the fused/unfused cache update implementations.
         self.attn.kv_cache = kv_cache
 
         attn_metadata = self.builder.build(
@@ -253,7 +244,7 @@ def _run_qk_norm_rope_kvcache_fusion_test(
     block_size: int,
     is_neox: bool,
     use_shuffle_kv_layout: str,
-    kv_stride_order: tuple[int, ...],
+    kv_layout: KVCacheLayout,
     dtype: torch.dtype,
     kv_cache_dtype: str,
     rms_norm_eps: float,
@@ -326,7 +317,7 @@ def _run_qk_norm_rope_kvcache_fusion_test(
         # Run unfused (eager) forward
         with set_forward_context(None, vllm_config):
             forward_context = get_forward_context()
-            attn_metadata = model.build_attn_metadata(num_tokens, kv_stride_order)
+            attn_metadata = model.build_attn_metadata(num_tokens, kv_layout)
             forward_context.slot_mapping = {
                 model.layer_name: attn_metadata.slot_mapping
             }
@@ -341,7 +332,7 @@ def _run_qk_norm_rope_kvcache_fusion_test(
         with set_forward_context(None, vllm_config):
             model_fused = torch.compile(model, backend=backend)
             forward_context = get_forward_context()
-            attn_metadata = model_fused.build_attn_metadata(num_tokens, kv_stride_order)
+            attn_metadata = model_fused.build_attn_metadata(num_tokens, kv_layout)
             forward_context.slot_mapping = {
                 model.layer_name: attn_metadata.slot_mapping
             }
@@ -419,10 +410,10 @@ _FUSION_CONFIGS = [
 @pytest.mark.parametrize("num_tokens", [5, 16, 2048])
 @pytest.mark.parametrize("use_shuffle_kv_layout", ["1", "0"])
 @pytest.mark.parametrize(
-    "kv_stride_order",
+    "kv_layout",
     [
-        pytest.param((0, 1, 2, 3, 4), id="block_first"),
-        pytest.param((1, 0, 2, 3, 4), id="kv_first"),
+        pytest.param(KVCacheLayout.LBHNC, id="head_major"),
+        pytest.param(KVCacheLayout.LBNHC, id="token_major"),
     ],
 )
 @pytest.mark.parametrize("enable_aiter_triton_rope", [True, False])
@@ -435,6 +426,7 @@ _FUSION_CONFIGS = [
     not is_aiter_found_and_supported(),
     reason="Only test on ROCm with AITER installed and supported",
 )
+@pytest.mark.skip(reason="AITER fusion does not support packed standardized K/V caches")
 def test_qk_norm_rope_kvcache_fusion(
     num_tokens: int,
     num_heads: int,
@@ -445,7 +437,7 @@ def test_qk_norm_rope_kvcache_fusion(
     attn_backend: AttentionBackendEnum,
     enable_aiter_triton_rope: bool,
     use_shuffle_kv_layout: str,
-    kv_stride_order: tuple[int, ...],
+    kv_layout: KVCacheLayout,
     block_size: int,
     dtype: torch.dtype,
     kv_cache_dtype: str,
@@ -469,7 +461,7 @@ def test_qk_norm_rope_kvcache_fusion(
         block_size=block_size,
         is_neox=is_neox,
         use_shuffle_kv_layout=use_shuffle_kv_layout,
-        kv_stride_order=kv_stride_order,
+        kv_layout=kv_layout,
         dtype=dtype,
         kv_cache_dtype=kv_cache_dtype,
         rms_norm_eps=rms_norm_eps,

tests/compile/passes/test_rope_kvcache_fusion.py

@@ -37,6 +37,10 @@ from vllm.v1.attention.backend import (
     CommonAttentionMetadata,
 )
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
+from vllm.v1.kv_cache_interface import (
+    FullAttentionSpec,
+    compute_layer_kv_cache_shape_bytes,
+)
 
 INDEX_SELECT_OP = torch.ops.aten.index.Tensor
 VLLM_UNIFIED_KV_CACHE_UPDATE_OP = torch.ops.vllm.unified_kv_cache_update
@@ -136,28 +140,21 @@ class QKRoPEKVCacheTestModel(torch.nn.Module):
         max_blocks = (max(batch_spec.seq_lens) + self.block_size - 1) // self.block_size
         num_blocks = batch_size * max_blocks
 
-        # Fetch the attention backend and kv cache shape and stride order
-        kv_cache_shape = self.attn_backend.get_kv_cache_shape(
-            num_blocks, self.block_size, self.num_kv_heads, self.head_size
+        kv_cache_shape = compute_layer_kv_cache_shape_bytes(
+            FullAttentionSpec(
+                block_size=self.block_size,
+                num_kv_heads=self.num_kv_heads,
+                head_size=self.head_size,
+                dtype=self.kv_cache_dtype,
+            ),
+            num_blocks,
         )
-        try:
-            kv_cache_stride_order = self.attn_backend.get_kv_cache_stride_order()
-        except (AttributeError, NotImplementedError):
-            kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
 
-        kv_cache_shape = tuple(kv_cache_shape[i] for i in kv_cache_stride_order)
-        inv_order = [
-            kv_cache_stride_order.index(i) for i in range(len(kv_cache_stride_order))
-        ]
-
-        # Create dummy KV cache
-        raw_tensor = torch.zeros(
-            2 * num_blocks * self.block_size * self.num_kv_heads * self.head_size,
-            dtype=self.kv_cache_dtype,
+        kv_cache = torch.zeros(
+            kv_cache_shape,
+            dtype=torch.int8,
             device=self.device,
-        )
-        raw_tensor = raw_tensor.view(kv_cache_shape)
-        kv_cache = raw_tensor.permute(*inv_order)
+        ).view(self.kv_cache_dtype)
 
         self.attn.kv_cache = kv_cache
 

tests/distributed/test_kvlayout.py

@@ -21,7 +21,7 @@ def test_get_kv_connector_cache_layout_without_kv_connector():
     with set_current_vllm_config(vllm_config):
         # Test with default settings
         layout = get_kv_connector_cache_layout()
-        assert layout == "NHD"
+        assert layout is None
 
 
 def test_get_kv_connector_cache_layout_with_lmcache_connector():
@@ -35,7 +35,7 @@ def test_get_kv_connector_cache_layout_with_lmcache_connector():
     with set_current_vllm_config(vllm_config):
         # Test with default settings
         layout = get_kv_connector_cache_layout()
-        assert layout == "NHD"
+        assert layout is None
 
 
 def test_get_kv_connector_cache_layout_with_nixl_connector():
@@ -52,7 +52,7 @@ def test_get_kv_connector_cache_layout_with_nixl_connector():
     with set_current_vllm_config(vllm_config):
         # Test with default settings
         layout = get_kv_connector_cache_layout()
-        assert layout == "HND"
+        assert layout == "LBHNC"
 
 
 def test_get_kv_connector_cache_layout_with_multi_connector():
@@ -75,4 +75,4 @@ def test_get_kv_connector_cache_layout_with_multi_connector():
     with set_current_vllm_config(vllm_config):
         # Test with default settings
         layout = get_kv_connector_cache_layout()
-        assert layout == "HND"
+        assert layout == "LBHNC"

tests/distributed/test_rocm_quick_reduce.py

@@ -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)