fused padding to kernel

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Add DCP + Eagle support for Tokenspeed MLA backends (#48180 )
2026-07-13 22:13:11 +00:00 · 2026-07-13 11:46:02 -07:00 · 2026-07-13 18:38:10 +00:00 · 2026-07-13 14:06:10 -04:00 · 2026-07-13 17:43:59 +00:00 · 2026-07-13 17:28:50 +00:00
362 changed files with 6369 additions and 1090 deletions
@@ -18,7 +18,7 @@ steps:
  - label: "XPU example Test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 50
    optional: true
    device: intel_gpu
    agent_tags:
@@ -39,7 +39,7 @@ steps:
  - label: "XPU V1 test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 70
    optional: true
    device: intel_gpu
    agent_tags:
@@ -60,7 +60,7 @@ steps:
  - label: "XPU server test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 45
    optional: true
    device: intel_gpu
    agent_tags:
@@ -3,7 +3,7 @@ depends_on:
  - image-build-xpu
 steps:
 - label: XPU Sleep Mode
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
    label: production
@@ -86,7 +86,7 @@ steps:
      pytest -v -s lora/test_punica_ops.py::test_add_lora_fused_moe_early_exit'

 - label: LoRA Punica FP8/XPU Ops
-  timeout_in_minutes: 45
+  timeout_in_minutes: 60
  device: intel_gpu
  agent_tags:
    label: production
@@ -3,7 +3,7 @@ depends_on:
  - image-build-xpu
 steps:
 - label: V1 Core + KV + Metrics
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
    label: production
@@ -33,7 +33,7 @@ steps:
      pytest -v -s v1/executor'

 - label: V1 Sample + Logits
-  timeout_in_minutes: 30
+  timeout_in_minutes: 90
  device: intel_gpu
  agent_tags:
    label: production
@@ -152,7 +152,7 @@ steps:

 - label: Regression
  key: regression
-  timeout_in_minutes: 30
+  timeout_in_minutes: 50
  device: intel_gpu
  agent_tags:
    label: production
@@ -186,7 +186,7 @@ steps:

 - label: Metrics, Tracing (2 GPUs)
  key: metrics-tracing-2-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  num_devices: 2
  device: intel_gpu
  agent_tags:
@@ -222,7 +222,7 @@ steps:

 - label: Async Engine, Inputs, Utils, Worker
  key: async-engine-inputs-utils-worker
-  timeout_in_minutes: 30
+  timeout_in_minutes: 55
  device: intel_gpu
  agent_tags:
    label: production
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Distributed Model Tests (2 GPUs)
  key: distributed-model-tests-2-gpus
-  timeout_in_minutes: 50
+  timeout_in_minutes: 65
  device: intel_gpu
  agent_tags:
    label: production
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: "Multi-Modal Models (Standard) 1: qwen2"
  key: multi-modal-models-standard-1-qwen2
-  timeout_in_minutes: 45
+  timeout_in_minutes: 70
  device: intel_gpu
  agent_tags:
    label: production
@@ -29,7 +29,7 @@ steps:

 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
  key: multi-modal-models-standard-2-qwen3-gemma
-  timeout_in_minutes: 45
+  timeout_in_minutes: 70
  device: intel_gpu
  agent_tags:
    label: production
@@ -52,7 +52,7 @@ steps:

 - label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
  key: multi-modal-models-standard-3-llava-qwen2-vl
-  timeout_in_minutes: 45
+  timeout_in_minutes: 65
  device: intel_gpu
  agent_tags:
    label: production
@@ -100,7 +100,7 @@ steps:

 - label: Multi-Modal Processor # 44min
  key: multi-modal-processor
-  timeout_in_minutes: 45
+  timeout_in_minutes: 60
  device: intel_gpu
  agent_tags:
    label: production
@@ -17,7 +17,7 @@ steps:
  - label: "XPU example Test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 50
    device: intel_gpu
    agent_tags:
      label: production
@@ -76,7 +76,7 @@ steps:
  - label: "XPU V1 test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 70
    device: intel_gpu
    agent_tags:
      label: production
@@ -104,7 +104,7 @@ steps:
  - label: "XPU server test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 45
    device: intel_gpu
    agent_tags:
      label: production
@@ -15,9 +15,9 @@ set -euo pipefail

 DEFAULT_REPO_SLUG="vllm-project/vllm"
 DEFAULT_CI_HCL_SOURCE="docker/ci-rocm.hcl"
-DEFAULT_CI_BASE_CONTENT_FILES="requirements/common.txt requirements/rocm.txt requirements/test/rocm.txt docker/Dockerfile.rocm_base docker/ci-rocm.hcl docker/docker-bake-rocm.hcl tools/install_torchcodec_rocm.sh tests/vllm_test_utils .buildkite/scripts/ci-bake-rocm.sh .buildkite/scripts/rocm/build-ci-base.sh"
+DEFAULT_CI_BASE_CONTENT_FILES="requirements/common.txt requirements/rocm.txt requirements/test/rocm.txt docker/Dockerfile.rocm_base docker/ci-rocm.hcl docker/docker-bake-rocm.hcl tools/install_torchcodec_rocm.sh tools/install_protoc.sh rust-toolchain.toml tests/vllm_test_utils .buildkite/scripts/ci-bake-rocm.sh .buildkite/scripts/rocm/build-ci-base.sh"
 DEFAULT_CI_BASE_DOCKERFILE="docker/Dockerfile.rocm"
-DEFAULT_CI_BASE_DOCKERFILE_STAGES="base build_rixl build_rocshmem build_deepep mori_base ci_base"
+DEFAULT_CI_BASE_DOCKERFILE_STAGES="base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust-toolchain build_rixl build_rocshmem build_deepep mori_base ci_base"
 DEFAULT_CI_BASE_METADATA_VERSION="1"
 IMAGE_EXISTED_BEFORE_BUILD=0

@@ -764,7 +764,7 @@ configure_ci_base_image_refs() {
        fi
        set_buildkite_metadata "rocm-ci-base-image" "${CI_BASE_IMAGE_TAG}"
        set_buildkite_metadata "rocm-ci-base-image-content" "${content_tag}"
-        set_buildkite_metadata "rocm-ci-base-image-commit" "${CI_BASE_IMAGE_TAG_COMMIT:-}"
+        set_buildkite_metadata "rocm-ci-base-image-commit" "${CI_BASE_IMAGE_TAG_COMMIT_REF:-}"
        set_buildkite_metadata "rocm-ci-base-image-stable" "${CI_BASE_IMAGE_TAG_STABLE:-}"
        return 0
    fi
@@ -1211,12 +1211,24 @@ uses_rocm_csrc_cache() {
    esac
 }

+uses_rocm_rust_cache() {
+    case "${TARGET}" in
+        rust-rocm-ci|test-rocm-ci|test-rocm-ci-with-wheel|test-rocm-ci-with-artifacts|export-wheel-rocm)
+            return 0
+            ;;
+        *)
+            return 1
+            ;;
+    esac
+}
+
 compute_rocm_csrc_content_hash() {
    local bake_dir=""
    local dockerfile_rocm=""
    local -a content_paths=(
        "requirements/common.txt"
        "requirements/rocm.txt"
+        "pyproject.toml"
        "setup.py"
        "CMakeLists.txt"
        "cmake"
@@ -1260,6 +1272,56 @@ compute_rocm_csrc_content_hash_if_needed() {
    echo "ROCm csrc content cache ref: ${ROCM_CSRC_CONTENT_CACHE_REF}"
 }

+compute_rocm_rust_content_hash() {
+    local bake_dir=""
+    local dockerfile_rocm=""
+    local -a content_paths=(
+        "requirements/build/rust.txt"
+        "rust/Cargo.lock"
+        "rust/Cargo.toml"
+        "rust/proto"
+        "rust/src"
+        "rust-toolchain.toml"
+        "tools/build_rust.py"
+        "tools/install_protoc.sh"
+        "build_rust.sh"
+    )
+    local -a content_args=()
+
+    bake_dir=$(dirname "${VLLM_BAKE_FILE}")
+    dockerfile_rocm="${bake_dir}/Dockerfile.rocm"
+    mapfile -t content_args < <(
+        get_content_arg_names "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build" "${ROCM_RUST_CONTENT_ARGS:-}"
+    )
+
+    {
+        printf 'rust-input-files-hash:%s\n' "$(compute_content_hash "${content_paths[@]}")"
+        printf 'dockerfile:%s\n' "${dockerfile_rocm}"
+        printf 'resolved-build-args:\n'
+        hash_dockerfile_arg_values "${dockerfile_rocm}" "${content_args[@]}"
+        printf 'dockerfile-stages:base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build\n'
+        if [[ -f "${dockerfile_rocm}" ]]; then
+            hash_dockerfile_stages "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build"
+        else
+            printf 'missing:%s\n' "${dockerfile_rocm}"
+        fi
+    } | sha256sum | cut -d' ' -f1
+}
+
+compute_rocm_rust_content_hash_if_needed() {
+    local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
+
+    if [[ "${ROCM_RUST_CONTENT_CACHE:-1}" == "0" ]] || ! uses_rocm_rust_cache; then
+        return 0
+    fi
+
+    ROCM_RUST_CONTENT_HASH=$(compute_rocm_rust_content_hash)
+    ROCM_RUST_CONTENT_CACHE_REF="${cache_repo}:rust-rocm-input-${ROCM_RUST_CONTENT_HASH}"
+    export ROCM_RUST_CONTENT_HASH
+    export ROCM_RUST_CONTENT_CACHE_REF
+    echo "ROCm Rust content cache ref: ${ROCM_RUST_CONTENT_CACHE_REF}"
+}
+
 write_hcl_string_list_entries() {
    local indent="$1"
    local value=""
@@ -1317,6 +1379,7 @@ write_rocm_build_arg_override() {
                "${CI_BASE_DOCKERFILE_STAGES:-${DEFAULT_CI_BASE_DOCKERFILE_STAGES}}" \
                "${CI_BASE_CONTENT_ARGS:-}"
            get_content_arg_names "${dockerfile_rocm}" "base csrc-build" "${ROCM_CSRC_CONTENT_ARGS:-}"
+            get_content_arg_names "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build" "${ROCM_RUST_CONTENT_ARGS:-}"
        } | awk 'NF && !seen[$0]++'
    )

@@ -1365,46 +1428,133 @@ validate_cache_export_mode() {
    esac
 }

+validate_content_cache_export_mode() {
+    local mode="$1"
+    local env_name="$2"
+
+    case "${mode}" in
+        missing|always|never)
+            ;;
+        *)
+            echo "Error: ${env_name} must be one of: missing, always, never"
+            exit 1
+            ;;
+    esac
+}
+
+should_export_content_cache_ref() {
+    local cache_ref="$1"
+    local cache_name="$2"
+    local mode="${ROCM_CONTENT_CACHE_EXPORT_MODE:-missing}"
+
+    case "${mode}" in
+        always)
+            echo "${cache_name} content cache export mode is always; exporting ${cache_ref}"
+            return 0
+            ;;
+        never)
+            echo "${cache_name} content cache export mode is never; not exporting ${cache_ref}"
+            return 1
+            ;;
+        missing|"")
+            if docker buildx imagetools inspect "${cache_ref}" >/dev/null 2>&1; then
+                echo "${cache_name} content cache exists; not re-exporting ${cache_ref}"
+                return 1
+            fi
+            echo "${cache_name} content cache missing; will export ${cache_ref}"
+            return 0
+            ;;
+        *)
+            echo "Error: ROCM_CONTENT_CACHE_EXPORT_MODE must be one of: missing, always, never"
+            exit 1
+            ;;
+    esac
+}
+
 write_rocm_cache_override() {
    local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
+    local content_cache_export_mode="${ROCM_CONTENT_CACHE_EXPORT_MODE:-missing}"
    local csrc_cache_to_mode="${ROCM_CSRC_CACHE_TO_MODE:-max}"
+    local rust_cache_to_mode="${ROCM_RUST_CACHE_TO_MODE:-max}"
    local rocm_cache_to_mode="${ROCM_FINAL_CACHE_TO_MODE:-min}"
-    local -a content_cache_from=()
+    local -a csrc_content_cache_from=()
+    local -a rust_content_cache_from=()
+    local -a combined_content_cache_from=()
    local -a csrc_cache_to=()
+    local -a rust_cache_to=()
    local -a rocm_cache_to=()
    local -a export_wheel_cache_to=()
+    local export_csrc_cache=1
+    local export_rust_cache=1

-    if ! uses_rocm_csrc_cache; then
+    if ! uses_rocm_csrc_cache && ! uses_rocm_rust_cache; then
        return 0
    fi

+    validate_content_cache_export_mode \
+        "${content_cache_export_mode}" \
+        "ROCM_CONTENT_CACHE_EXPORT_MODE"
    validate_cache_export_mode "${csrc_cache_to_mode}" "ROCM_CSRC_CACHE_TO_MODE"
+    validate_cache_export_mode "${rust_cache_to_mode}" "ROCM_RUST_CACHE_TO_MODE"
    validate_cache_export_mode "${rocm_cache_to_mode}" "ROCM_FINAL_CACHE_TO_MODE"
+    echo "ROCm content cache export mode: ${content_cache_export_mode}"
    echo "ROCm csrc cache export mode: ${csrc_cache_to_mode}"
+    echo "ROCm Rust cache export mode: ${rust_cache_to_mode}"
    echo "ROCm final image cache export mode: ${rocm_cache_to_mode}"

    if [[ -n "${ROCM_CSRC_CONTENT_CACHE_REF:-}" ]]; then
-        content_cache_from+=("type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF}")
-        csrc_cache_to+=(
-            "type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF},mode=${csrc_cache_to_mode},ignore-error=true"
-        )
+        csrc_content_cache_from+=("type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF}")
+        if should_export_content_cache_ref "${ROCM_CSRC_CONTENT_CACHE_REF}" "ROCm csrc"; then
+            csrc_cache_to+=(
+                "type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF},mode=${csrc_cache_to_mode},ignore-error=true"
+            )
+        else
+            export_csrc_cache=0
+        fi
    fi

+    if [[ -n "${ROCM_RUST_CONTENT_CACHE_REF:-}" ]]; then
+        rust_content_cache_from+=("type=registry,ref=${ROCM_RUST_CONTENT_CACHE_REF}")
+        if should_export_content_cache_ref "${ROCM_RUST_CONTENT_CACHE_REF}" "ROCm Rust"; then
+            rust_cache_to+=(
+                "type=registry,ref=${ROCM_RUST_CONTENT_CACHE_REF},mode=${rust_cache_to_mode},ignore-error=true"
+            )
+        else
+            export_rust_cache=0
+        fi
+    fi
+
+    combined_content_cache_from=("${csrc_content_cache_from[@]}" "${rust_content_cache_from[@]}")
+
    # Docker Hub cache exports are best-effort. A cache-only target failure can
    # otherwise cancel the sibling image target before its manifest is pushed.
    if [[ -n "${BUILDKITE_COMMIT:-}" ]]; then
-        csrc_cache_to+=(
-            "type=registry,ref=${cache_repo}:csrc-rocm-${BUILDKITE_COMMIT},mode=${csrc_cache_to_mode},ignore-error=true"
-        )
+        if [[ ${export_csrc_cache} -eq 1 ]]; then
+            csrc_cache_to+=(
+                "type=registry,ref=${cache_repo}:csrc-rocm-${BUILDKITE_COMMIT},mode=${csrc_cache_to_mode},ignore-error=true"
+            )
+        fi
+        if [[ ${export_rust_cache} -eq 1 ]]; then
+            rust_cache_to+=(
+                "type=registry,ref=${cache_repo}:rust-rocm-${BUILDKITE_COMMIT},mode=${rust_cache_to_mode},ignore-error=true"
+            )
+        fi
        rocm_cache_to+=(
            "type=registry,ref=${cache_repo}:rocm-${BUILDKITE_COMMIT},mode=${rocm_cache_to_mode},ignore-error=true"
        )
    fi

    if [[ -n "${ROCM_CACHE_BRANCH_TAG:-}" ]]; then
-        csrc_cache_to+=(
-            "type=registry,ref=${cache_repo}:csrc-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${csrc_cache_to_mode},ignore-error=true"
-        )
+        if [[ ${export_csrc_cache} -eq 1 ]]; then
+            csrc_cache_to+=(
+                "type=registry,ref=${cache_repo}:csrc-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${csrc_cache_to_mode},ignore-error=true"
+            )
+        fi
+        if [[ ${export_rust_cache} -eq 1 ]]; then
+            rust_cache_to+=(
+                "type=registry,ref=${cache_repo}:rust-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${rust_cache_to_mode},ignore-error=true"
+            )
+        fi
        rocm_cache_to+=(
            "type=registry,ref=${cache_repo}:rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${rocm_cache_to_mode},ignore-error=true"
        )
@@ -1422,7 +1572,7 @@ target "csrc-rocm-ci" {
  cache-from = concat(
    get_cache_from_rocm_csrc(),
 EOF
-        write_hcl_string_list "    " "${content_cache_from[@]}"
+        write_hcl_string_list "    " "${csrc_content_cache_from[@]}"
        cat <<EOF
  )
 EOF
@@ -1430,11 +1580,23 @@ EOF
        cat <<EOF
 }

+target "rust-rocm-ci" {
+  cache-from = concat(
+    get_cache_from_rocm_rust(),
+EOF
+        write_hcl_string_list "    " "${rust_content_cache_from[@]}"
+        cat <<EOF
+  )
+EOF
+        write_hcl_string_list_attr "  " "cache-to" "${rust_cache_to[@]}"
+        cat <<EOF
+}
+
 target "test-rocm-ci" {
  cache-from = concat(
    get_cache_from_rocm(),
 EOF
-        write_hcl_string_list "    " "${content_cache_from[@]}"
+        write_hcl_string_list "    " "${combined_content_cache_from[@]}"
        cat <<EOF
  )
 EOF
@@ -1446,7 +1608,7 @@ target "export-wheel-rocm" {
  cache-from = concat(
    get_cache_from_rocm(),
 EOF
-        write_hcl_string_list "    " "${content_cache_from[@]}"
+        write_hcl_string_list "    " "${combined_content_cache_from[@]}"
        cat <<EOF
  )
 EOF
@@ -1920,6 +2082,7 @@ main() {
    compute_dependency_cache_keys
    write_ci_base_label_override
    compute_rocm_csrc_content_hash_if_needed
+    compute_rocm_rust_content_hash_if_needed
    write_rocm_cache_override
    resolve_ci_base_dependency_targets
    print_bake_config
@@ -210,7 +210,7 @@ repos:
    name: Check SPDX headers
    entry: python tools/pre_commit/check_spdx_header.py
    language: python
-    types: [python]
+    types_or: [python, rust, proto]
  - id: check-root-lazy-imports
    name: Check root lazy imports
    entry: python tools/pre_commit/check_init_lazy_imports.py
@@ -1116,6 +1116,164 @@ void fused_sigmoid_gating_delta_rule_update_kernel_impl(
  });
 }

+// Speculative-decode variant: processes a varlen batch where each sequence has
+// ``q_len`` draft tokens, runs the recurrence sequentially over those tokens
+// (inside the kernel, so one dispatch handles the whole draft block), reads the
+// initial state from cache slot ``num_accepted-1`` and stores the state *after*
+// token ``t`` into cache slot ``t`` (multi-slot rollback, matching the GPU
+// kernel). Parallelized over (sequence, v_head); the per-sequence token loop is
+// sequential as required by the recurrence.
+template <typename scalar_t, typename param_t>
+void fused_sigmoid_gating_delta_rule_update_spec_kernel_impl(
+    const scalar_t* __restrict__ q_ptr,  // [T, HK, EK]
+    const scalar_t* __restrict__ k_ptr,  // [T, HK, EK]
+    const scalar_t* __restrict__ v_ptr,  // [T, HV, EV]
+    const param_t* __restrict__ A_log_ptr,
+    const scalar_t* __restrict__ a_ptr,  // [T, HV]
+    const scalar_t* __restrict__ dt_bias_ptr,
+    const scalar_t* __restrict__ b_ptr,  // [T, HV]
+    const int32_t* __restrict__ spec_indices_ptr,  // [N, S]
+    const int32_t* __restrict__ num_accepted_ptr,  // [N]
+    const int32_t* __restrict__ cu_seqlens_ptr,     // [N + 1]
+    float* __restrict__ state_ptr,
+    scalar_t* __restrict__ o_ptr,  // [T, HV, EV]
+    float* __restrict__ qk_scale_buf,  // [2, T, HK]
+    int64_t total_tokens,
+    int64_t batch_size,
+    int64_t spec_stride,
+    int64_t num_heads,
+    int64_t head_dim,
+    int64_t v_num_heads,
+    int64_t v_head_dim,
+    int64_t q_strideT,
+    int64_t q_strideH,
+    int64_t k_strideT,
+    int64_t k_strideH,
+    int64_t v_strideT,
+    int64_t v_strideH,
+    int64_t state_slot_stride,
+    bool use_qk_l2norm_in_kernel,
+    double softplus_threshold) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int64_t VecSize = bVec::size();
+  constexpr int64_t fVecSize = fVec::size();
+  int64_t group_size = v_num_heads / num_heads;
+  double scale = 1 / std::sqrt((double)head_dim);
+  fVec scale_vec = fVec((float)scale);
+
+  if (use_qk_l2norm_in_kernel) {
+    float eps = 1e-5f;
+    at::parallel_for(0, total_tokens * num_heads, 0, [&](int64_t begin, int64_t end) {
+      for (int64_t i = begin; i < end; ++i) {
+        int64_t ti = i / num_heads;
+        int64_t ni = i % num_heads;
+        const scalar_t* qp = q_ptr + ti * q_strideT + ni * q_strideH;
+        const scalar_t* kp = k_ptr + ti * k_strideT + ni * k_strideH;
+        float sq = 0.f, sk = 0.f;
+        for (int64_t d = 0; d < head_dim; ++d) {
+          float qv = (float)qp[d];
+          sq += qv * qv;
+          float kv = (float)kp[d];
+          sk += kv * kv;
+        }
+        qk_scale_buf[ti * num_heads + ni] = 1.f / std::sqrt(sq + eps);
+        qk_scale_buf[total_tokens * num_heads + ti * num_heads + ni] = 1.f / std::sqrt(sk + eps);
+      }
+    });
+  }
+
+  at::parallel_for(0, batch_size * v_num_heads, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t idx = begin; idx < end; ++idx) {
+      int64_t bi = idx / v_num_heads;
+      int64_t ni = idx % v_num_heads;
+      int64_t kh = ni / group_size;
+      int64_t q_start = cu_seqlens_ptr[bi];
+      int64_t q_len = cu_seqlens_ptr[bi + 1] - q_start;
+      if (q_len <= 0) {
+        continue;
+      }
+      int64_t acc = (int64_t)num_accepted_ptr[bi];
+      // Clamp acc-1 to >=0: when num_accepted is 0 the unclamped index reads
+      // out of bounds and yields an arbitrary prev_slot used to index the SSM
+      // state. Mirrors the GPU guard tl.maximum(num_accepted - 1, 0).
+      int64_t prev_slot =
+          (int64_t)spec_indices_ptr[bi * spec_stride + (acc > 0 ? acc - 1 : 0)];
+      for (int64_t t = 0; t < q_len; ++t) {
+        int64_t cur_slot = (int64_t)spec_indices_ptr[bi * spec_stride + t];
+        int64_t token = q_start + t;
+        const float* src = state_ptr + prev_slot * state_slot_stride + ni * head_dim * v_head_dim;
+        float* dst = state_ptr + cur_slot * state_slot_stride + ni * head_dim * v_head_dim;
+        float g_val = -std::exp((float)A_log_ptr[ni]) *
+            softplus((float)a_ptr[token * v_num_heads + ni] + (float)dt_bias_ptr[ni], softplus_threshold);
+        float g_val_exp = std::exp(g_val);
+        fVec g_val_exp_vec = fVec(g_val_exp);
+        float beta_val = 1.f / (1.f + std::exp(-(float)b_ptr[token * v_num_heads + ni]));
+        fVec beta_vec = fVec(beta_val);
+        int64_t q_offset = token * q_strideT + kh * q_strideH;
+        int64_t k_offset = token * k_strideT + kh * k_strideH;
+        float q_scale = use_qk_l2norm_in_kernel ? qk_scale_buf[token * num_heads + kh] : 1.f;
+        float k_scale =
+            use_qk_l2norm_in_kernel ? qk_scale_buf[total_tokens * num_heads + token * num_heads + kh] : 1.f;
+        int64_t v_offset = token * v_strideT + ni * v_strideH;
+        int64_t o_offset = (token * v_num_heads + ni) * v_head_dim;
+        int64_t dvi = 0;
+        for (; dvi <= v_head_dim - VecSize; dvi += VecSize) {
+          fVec kv_mem_vec0 = fVec(0.f);
+          fVec kv_mem_vec1 = fVec(0.f);
+          for (int di = 0; di < head_dim; ++di) {
+            fVec k_val_vec = fVec((float)k_ptr[k_offset + di] * k_scale);
+            fVec sv0 = fVec::loadu(src + di * v_head_dim + dvi);
+            fVec sv1 = fVec::loadu(src + di * v_head_dim + dvi + fVecSize);
+            kv_mem_vec0 = kv_mem_vec0 + sv0 * g_val_exp_vec * k_val_vec;
+            kv_mem_vec1 = kv_mem_vec1 + sv1 * g_val_exp_vec * k_val_vec;
+          }
+          bVec v_bvec = bVec::loadu(v_ptr + v_offset + dvi);
+          fVec v_vec0, v_vec1;
+          std::tie(v_vec0, v_vec1) = at::vec::convert_to_float(v_bvec);
+          fVec dt_vec0 = (v_vec0 - kv_mem_vec0) * beta_vec;
+          fVec dt_vec1 = (v_vec1 - kv_mem_vec1) * beta_vec;
+          fVec o_vec0 = fVec(0.f);
+          fVec o_vec1 = fVec(0.f);
+          for (int di = 0; di < head_dim; ++di) {
+            fVec q_vec = fVec((float)q_ptr[q_offset + di] * q_scale);
+            fVec k_vec = fVec((float)k_ptr[k_offset + di] * k_scale);
+            fVec sv0 = fVec::loadu(src + di * v_head_dim + dvi);
+            fVec sv1 = fVec::loadu(src + di * v_head_dim + dvi + fVecSize);
+            sv0 = sv0 * g_val_exp_vec + k_vec * dt_vec0;
+            sv1 = sv1 * g_val_exp_vec + k_vec * dt_vec1;
+            o_vec0 = o_vec0 + sv0 * q_vec * scale_vec;
+            o_vec1 = o_vec1 + sv1 * q_vec * scale_vec;
+            sv0.store(dst + di * v_head_dim + dvi);
+            sv1.store(dst + di * v_head_dim + dvi + fVecSize);
+          }
+          bVec o_vec = at::vec::convert_from_float<scalar_t>(o_vec0, o_vec1);
+          o_vec.store(o_ptr + o_offset + dvi);
+        }
+        for (; dvi < v_head_dim; ++dvi) {
+          float kv_mem_val = 0.f;
+          for (int di = 0; di < head_dim; ++di) {
+            float k_val = (float)k_ptr[k_offset + di] * k_scale;
+            kv_mem_val += src[di * v_head_dim + dvi] * g_val_exp * k_val;
+          }
+          float v_val = (float)v_ptr[v_offset + dvi];
+          float dt_val = (v_val - kv_mem_val) * beta_val;
+          float o_val = 0.f;
+          for (int di = 0; di < head_dim; ++di) {
+            float q_val = (float)q_ptr[q_offset + di] * q_scale;
+            float k_val = (float)k_ptr[k_offset + di] * k_scale;
+            float ns = src[di * v_head_dim + dvi] * g_val_exp + k_val * dt_val;
+            dst[di * v_head_dim + dvi] = ns;
+            o_val += ns * q_val * scale;
+          }
+          o_ptr[o_offset + dvi] = (scalar_t)o_val;
+        }
+        prev_slot = cur_slot;
+      }
+    }
+  });
+}
+
 template <typename scalar_t>
 void fused_gdn_gating_kernel_impl(
    float* __restrict__ A_log,
@@ -1500,6 +1658,103 @@ at::Tensor fused_sigmoid_gating_delta_rule_update_cpu(
  return core_attn_out;
 }

+// Speculative-decode update (multi-token, multi-slot rollback).
+// q: [T, HK, EK]  k: [T, HK, EK]  v: [T, HV, EV]
+// a: [T, HV]  b: [T, HV]
+// initial_state_source: [N_slots, HV, EK, EV] FP32 (updated in place)
+// spec_state_indices: [batch, S] INT32 (S = num_spec + 1)
+// num_accepted_tokens: [batch] INT32
+// cu_seqlens: [batch + 1] INT32
+// Returns output: [T, HV, EV]
+at::Tensor fused_sigmoid_gating_delta_rule_update_spec_cpu(
+    const at::Tensor& A_log,
+    const at::Tensor& dt_bias,
+    const at::Tensor& q,
+    const at::Tensor& k,
+    const at::Tensor& v,
+    const at::Tensor& a,
+    const at::Tensor& b,
+    at::Tensor& initial_state_source,
+    const at::Tensor& spec_state_indices,
+    const at::Tensor& num_accepted_tokens,
+    const at::Tensor& cu_seqlens,
+    bool use_qk_l2norm_in_kernel,
+    double softplus_beta = 1.0,
+    double softplus_threshold = 20.0) {
+  CHECK_DIM(3, q);
+  CHECK_DIM(3, v);
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(q);
+  int64_t total_tokens = q.size(0);
+  int64_t num_heads = q.size(1);
+  int64_t head_dim = q.size(2);
+  int64_t v_num_heads = v.size(1);
+  int64_t v_head_dim = v.size(2);
+  int64_t batch_size = cu_seqlens.size(0) - 1;
+  int64_t spec_stride = spec_state_indices.stride(0);
+  CHECK_INPUT_SHAPE_DTYPE<true>(k, {total_tokens, num_heads, head_dim}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(v, {total_tokens, v_num_heads, v_head_dim}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(a, {total_tokens, v_num_heads}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(b, {total_tokens, v_num_heads}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(dt_bias, {v_num_heads}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(num_accepted_tokens, {batch_size}, at::kInt);
+  CHECK_INPUT_SHAPE_DTYPE<true>(cu_seqlens, {batch_size + 1}, at::kInt);
+  CHECK_EQ(v_num_heads % num_heads, 0);
+  TORCH_CHECK(A_log.sizes() == at::IntArrayRef({v_num_heads}));
+  CHECK_INPUT_SHAPE_DTYPE<true>(
+      initial_state_source,
+      {initial_state_source.size(0), v_num_heads, head_dim, v_head_dim},
+      at::kFloat);
+  TORCH_CHECK(initial_state_source.size(0) >= batch_size,
+      "initial_state_source capacity too small: size(0)=",
+      initial_state_source.size(0), ", batch_size=", batch_size);
+
+  int64_t q_strideT = q.stride(0);
+  int64_t q_strideH = q.stride(1);
+  int64_t k_strideT = k.stride(0);
+  int64_t k_strideH = k.stride(1);
+  int64_t v_strideT = v.stride(0);
+  int64_t v_strideH = v.stride(1);
+  int64_t state_slot_stride = initial_state_source.stride(0);
+
+  at::Tensor o = at::empty({total_tokens, v_num_heads, v_head_dim}, q.options());
+  at::Tensor qk_scale_buf = at::empty({2, total_tokens, num_heads}, at::kFloat);
+
+  CPU_DISPATCH_REDUCED_FLOATING_TYPES_EXT(
+      q.scalar_type(), A_log.scalar_type(), "fused_sigmoid_gating_delta_rule_update_spec_kernel_impl", [&] {
+        fused_sigmoid_gating_delta_rule_update_spec_kernel_impl<scalar_t, param_t>(
+            q.data_ptr<scalar_t>(),
+            k.data_ptr<scalar_t>(),
+            v.data_ptr<scalar_t>(),
+            A_log.data_ptr<param_t>(),
+            a.data_ptr<scalar_t>(),
+            dt_bias.data_ptr<scalar_t>(),
+            b.data_ptr<scalar_t>(),
+            spec_state_indices.data_ptr<int32_t>(),
+            num_accepted_tokens.data_ptr<int32_t>(),
+            cu_seqlens.data_ptr<int32_t>(),
+            initial_state_source.data_ptr<float>(),
+            o.data_ptr<scalar_t>(),
+            qk_scale_buf.data_ptr<float>(),
+            total_tokens,
+            batch_size,
+            spec_stride,
+            num_heads,
+            head_dim,
+            v_num_heads,
+            v_head_dim,
+            q_strideT,
+            q_strideH,
+            k_strideT,
+            k_strideH,
+            v_strideT,
+            v_strideH,
+            state_slot_stride,
+            use_qk_l2norm_in_kernel,
+            softplus_threshold);
+      });
+  return o;
+}
+
 // A_log: [num_v_heads]
 // a: [batch, num_v_heads]
 // b: [batch, num_v_heads]
@@ -120,6 +120,14 @@ at::Tensor fused_sigmoid_gating_delta_rule_update_cpu(
    bool use_qk_l2norm_in_kernel, double softplus_beta = 1.0,
    double softplus_threshold = 20.0);

+at::Tensor fused_sigmoid_gating_delta_rule_update_spec_cpu(
+    const at::Tensor& A_log, const at::Tensor& dt_bias, const at::Tensor& q,
+    const at::Tensor& k, const at::Tensor& v, const at::Tensor& a,
+    const at::Tensor& b, at::Tensor& initial_state_source,
+    const at::Tensor& spec_state_indices, const at::Tensor& num_accepted_tokens,
+    const at::Tensor& cu_seqlens, bool use_qk_l2norm_in_kernel,
+    double softplus_beta = 1.0, double softplus_threshold = 20.0);
+
 std::tuple<at::Tensor, at::Tensor> fused_gdn_gating_cpu(
    const at::Tensor& A_log, const at::Tensor& a, const at::Tensor& b,
    const at::Tensor& dt_bias);
@@ -508,6 +516,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
      "softplus_threshold=20.0) -> Tensor");
  ops.impl("fused_sigmoid_gating_delta_rule_update_cpu", torch::kCPU,
           &fused_sigmoid_gating_delta_rule_update_cpu);
+  ops.def(
+      "fused_sigmoid_gating_delta_rule_update_spec_cpu(Tensor A_log, Tensor "
+      "dt_bias, Tensor q, Tensor k, Tensor v, Tensor a, Tensor b, "
+      "Tensor(a!) initial_state_source, Tensor spec_state_indices, "
+      "Tensor num_accepted_tokens, Tensor cu_seqlens, bool "
+      "use_qk_l2norm_in_kernel, float softplus_beta=1.0, float "
+      "softplus_threshold=20.0) -> Tensor");
+  ops.impl("fused_sigmoid_gating_delta_rule_update_spec_cpu", torch::kCPU,
+           &fused_sigmoid_gating_delta_rule_update_spec_cpu);
  ops.def(
      "fused_gdn_gating_cpu(Tensor A_log, Tensor a, Tensor b, Tensor dt_bias) "
      "-> (Tensor, Tensor)");
@@ -691,19 +691,20 @@ void launchFusedDeepseekV4QNormRopeKVRopeQuantInsert(
 //
 // Sibling to the FlashMLA kernel above, used by the FlashInfer V4 sparse-MLA
 // backend.  Differences from the legacy path:
-//   * No Q head padding — output Q layout matches the input num_heads_q.
+//   * Optional FP8 Q head padding, with zero-filled padding heads written
+//     directly from this fused kernel.
 //   * KV is written as a *contiguous* 512-wide row per token (token-strided),
 //     not the legacy UE8M0 paged layout with a separate scale tail.
 //   * Q/KV are stored either as bf16 or as per-tensor E4M3 FP8 (one global
 //     scale), selected by the STORE_Q_FP8 / STORE_KV_FP8 template flags.
 //
-// Grid: 1D, gridDim.x = ceil(num_tokens_full * (num_heads_q + 1) / warps).
-// Each warp handles one (token, slot): slot < num_heads_q → Q, slot ==
-// num_heads_q → KV.
+// Grid: 1D, gridDim.x = ceil(num_tokens_full * (num_heads_q_padded + 1) /
+// warps). Each warp handles one (token, slot): slot < num_heads_q → live Q,
+// num_heads_q <= slot < num_heads_q_padded → padded Q, final slot → KV.
 template <typename scalar_t_in, bool STORE_Q_FP8, bool STORE_KV_FP8>
 __global__ void fusedDeepseekV4FullCacheKernel(
    scalar_t_in* __restrict__ q_inout,          // [N, H, 512], in place (bf16)
-    uint8_t* __restrict__ q_fp8_out,            // [N, H, 512] fp8, optional
+    uint8_t* __restrict__ q_fp8_out,            // [N, H_out, 512] fp8, optional
    int64_t const q_fp8_stride0,                // elements (fp8 == bytes)
    int64_t const q_fp8_stride1,                // elements (fp8 == bytes)
    scalar_t_in const* __restrict__ kv_in,      // [N, 512] bf16
@@ -716,7 +717,8 @@ __global__ void fusedDeepseekV4FullCacheKernel(
    float const eps,
    int const num_tokens_full,      // = q.size(0) = kv.size(0)
    int const num_tokens_insert,    // = slot_mapping.size(0)
-    int const num_heads_q,          // H (no padding)
+    int const num_heads_q,          // real H
+    int const num_heads_q_padded,   // output/kernel H
    int const cache_block_size,     // tokens per cache block
    int64_t const kv_block_stride,  // bytes per cache block
    int64_t const kv_token_stride) {  // bytes per cache token
@@ -731,11 +733,12 @@ __global__ void fusedDeepseekV4FullCacheKernel(
    int const laneId = threadIdx.x % 32;
    int const globalWarpIdx = blockIdx.x * warpsPerBlock + warpId;

-    int const slotsPerToken = num_heads_q + 1;
+    int const slotsPerToken = num_heads_q_padded + 1;
    int const tokenIdx = globalWarpIdx / slotsPerToken;
    int const slotIdx = globalWarpIdx % slotsPerToken;
    if (tokenIdx >= num_tokens_full) return;
-    bool const isKV = (slotIdx == num_heads_q);
+    bool const isKV = (slotIdx == num_heads_q_padded);
+    bool const isQPad = (slotIdx >= num_heads_q) && !isKV;
    // KV branch: skip DP-padded tokens (no slot reserved for them).
    if (isKV && tokenIdx >= num_tokens_insert) return;

@@ -744,6 +747,19 @@ __global__ void fusedDeepseekV4FullCacheKernel(
 #endif

    int const dim_base = laneId * kElemsPerLane;  // in [0, 512) step 16
+
+    if (isQPad) {
+      // padded position zero_
+      uint4 const zero{};
+      if constexpr (STORE_Q_FP8) {
+        uint8_t* dst = q_fp8_out +
+                       static_cast<int64_t>(tokenIdx) * q_fp8_stride0 +
+                       static_cast<int64_t>(slotIdx) * q_fp8_stride1 + dim_base;
+        *reinterpret_cast<uint4*>(dst) = zero;
+      }
+      return;
+    }
+
    scalar_t_in const* src_ptr;
    if (isKV) {
      src_ptr = kv_in + static_cast<int64_t>(tokenIdx) * kHeadDim + dim_base;
@@ -842,7 +858,8 @@ __global__ void fusedDeepseekV4FullCacheKernel(
        }
        scalar_t_in* dst =
            q_inout +
-            (static_cast<int64_t>(tokenIdx) * num_heads_q + slotIdx) * kHeadDim +
+            (static_cast<int64_t>(tokenIdx) * num_heads_q + slotIdx) *
+                kHeadDim +
            dim_base;
        *reinterpret_cast<uint4*>(dst) = out0;
        *reinterpret_cast<uint4*>(dst + 8) = out1;
@@ -898,12 +915,13 @@ static void launchFullCacheKernel(
    float const* cos_sin_cache, float const* fp8_scale,
    float const* q_fp8_scale_inv, float const eps, int const num_tokens_full,
    int const num_tokens_insert, int const num_heads_q,
-    int const cache_block_size, int64_t const kv_block_stride,
-    int64_t const kv_token_stride, char const* op_name, cudaStream_t stream) {
+    int const num_heads_q_padded, int const cache_block_size,
+    int64_t const kv_block_stride, int64_t const kv_token_stride,
+    char const* op_name, cudaStream_t stream) {
  constexpr int kBlockSize = 256;
  constexpr int kWarpsPerBlock = kBlockSize / 32;
  int64_t const total_warps =
-      static_cast<int64_t>(num_tokens_full) * (num_heads_q + 1);
+      static_cast<int64_t>(num_tokens_full) * (num_heads_q_padded + 1);
  int const grid =
      static_cast<int>((total_warps + kWarpsPerBlock - 1) / kWarpsPerBlock);
  auto* kernel =
@@ -926,13 +944,14 @@ static void launchFullCacheKernel(
                     q_fp8_stride1, kv_in, k_cache, slot_mapping, position_ids,
                     cos_sin_cache, fp8_scale, q_fp8_scale_inv, eps,
                     num_tokens_full, num_tokens_insert, num_heads_q,
-                     cache_block_size, kv_block_stride, kv_token_stride);
+                     num_heads_q_padded, cache_block_size, kv_block_stride,
+                     kv_token_stride);
 #else
  kernel<<<grid, kBlockSize, 0, stream>>>(
      q_inout, q_fp8_out, q_fp8_stride0, q_fp8_stride1, kv_in, k_cache,
-      slot_mapping, position_ids, cos_sin_cache, fp8_scale, q_fp8_scale_inv,
-      eps, num_tokens_full, num_tokens_insert, num_heads_q, cache_block_size,
-      kv_block_stride, kv_token_stride);
+      slot_mapping, position_ids, cos_sin_cache, fp8_scale,
+      q_fp8_scale_inv, eps, num_tokens_full, num_tokens_insert, num_heads_q,
+      num_heads_q_padded, cache_block_size, kv_block_stride, kv_token_stride);
 #endif
 }

@@ -1090,8 +1109,8 @@ void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert(
            position_ids.const_data_ptr<int64_t>(),
            cos_sin_cache.const_data_ptr<float>(), nullptr, nullptr,
            static_cast<float>(eps), num_tokens_full, num_tokens_insert,
-            num_heads_q, static_cast<int>(cache_block_size), kv_block_stride,
-            kv_token_stride,
+            num_heads_q, num_heads_q,
+            static_cast<int>(cache_block_size), kv_block_stride, kv_token_stride,
            "fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert",
            stream);
      });
@@ -1116,8 +1135,9 @@ void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert(
  STD_TORCH_CHECK(q_fp8.device().is_cuda() && q_fp8.is_contiguous() &&
                      q_fp8.scalar_type() == ScalarType::Float8_e4m3fn &&
                      q_fp8.dim() == 3 && q_fp8.size(0) == q.size(0) &&
-                      q_fp8.size(1) == q.size(1) && q_fp8.size(2) == q.size(2),
-                  "q_fp8 must be a contiguous float8_e4m3fn tensor matching q");
+                      q_fp8.size(1) >= q.size(1) && q_fp8.size(2) == q.size(2),
+                  "q_fp8 must be a contiguous float8_e4m3fn tensor with shape "
+                  "[N, H_out, 512] and H_out >= q.size(1)");
  STD_TORCH_CHECK(k_cache.device().is_cuda(), "k_cache must be CUDA");
  STD_TORCH_CHECK(slot_mapping.device().is_cuda() &&
                      slot_mapping.scalar_type() == ScalarType::Long,
@@ -1155,6 +1175,7 @@ void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert(
  STD_TORCH_CHECK(num_tokens_insert <= num_tokens_full,
                  "slot_mapping must not exceed q row count");
  int const num_heads_q = static_cast<int>(q.size(1));
+  int const num_heads_q_padded = static_cast<int>(q_fp8.size(1));

  const torch::stable::accelerator::DeviceGuard device_guard(
      q.get_device_index());
@@ -1167,8 +1188,7 @@ void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert(
                                                           true>(
            // q is read-only in the fp8 path (the kernel writes q_fp8); the
            // launcher signature is non-const, so cast away const on the ptr.
-            reinterpret_cast<scalar_t*>(
-                const_cast<void*>(q.const_data_ptr())),
+            reinterpret_cast<scalar_t*>(const_cast<void*>(q.const_data_ptr())),
            reinterpret_cast<uint8_t*>(q_fp8.mutable_data_ptr()),
            q_fp8.stride(0), q_fp8.stride(1),
            reinterpret_cast<scalar_t const*>(kv.const_data_ptr()),
@@ -1179,7 +1199,7 @@ void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert(
            fp8_scale.const_data_ptr<float>(),
            q_fp8_scale_inv.const_data_ptr<float>(), static_cast<float>(eps),
            num_tokens_full, num_tokens_insert, num_heads_q,
-            static_cast<int>(cache_block_size),
+            num_heads_q_padded, static_cast<int>(cache_block_size),
            // fp8 cache: 1 byte/element -> stride already in bytes.
            k_cache.stride(0), k_cache.stride(1),
            "fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert",
@@ -1,5 +1,7 @@
 # default base image
 ARG REMOTE_VLLM="0"
+ARG VLLM_REPO="https://github.com/vllm-project/vllm.git"
+ARG VLLM_BRANCH="main"
 ARG COMMON_WORKDIR=/app
 ARG BASE_IMAGE=rocm/vllm-dev:base
 ARG CI_BASE_IMAGE=rocm/vllm-dev:ci_base
@@ -110,10 +112,8 @@ WORKDIR ${COMMON_WORKDIR}
 FROM base AS fetch_vllm_0
 ONBUILD COPY ./ vllm/
 FROM base AS fetch_vllm_1
-ARG VLLM_REPO="https://github.com/vllm-project/vllm.git"
-ARG VLLM_BRANCH="main"
-ENV VLLM_REPO=${VLLM_REPO}
-ENV VLLM_BRANCH=${VLLM_BRANCH}
+ARG VLLM_REPO
+ARG VLLM_BRANCH
 ONBUILD RUN git clone ${VLLM_REPO} \
 	    && cd vllm \
 	    && git fetch -v --prune -- origin ${VLLM_BRANCH} \
@@ -124,34 +124,118 @@ ONBUILD RUN git clone ${VLLM_REPO} \
 FROM fetch_vllm_${REMOTE_VLLM} AS fetch_vllm

 # -----------------------
-# Rust build stage
-# Builds the `vllm-rs` frontend in a dedicated stage so the wheel build stages
-# don't need the rust toolchain or protoc.
-FROM fetch_vllm AS rust-build
-ARG COMMON_WORKDIR
-ARG USE_SCCACHE
+# Rust/protoc toolchain inputs
+#
+# Keep these separate from Rust source inputs so Rust source-only changes do not
+# invalidate the long-lived ci_base image that carries the toolchain.
+FROM scratch AS rust_toolchain_input_0
+COPY rust-toolchain.toml /rust-toolchain-input/vllm/rust-toolchain.toml
+COPY tools/install_protoc.sh /rust-toolchain-input/vllm/tools/install_protoc.sh
+
+FROM base AS rust_toolchain_input_1
+ARG VLLM_REPO
+ARG VLLM_BRANCH
+RUN git clone --no-checkout --filter=blob:none ${VLLM_REPO} /rust-toolchain-input/vllm \
+    && cd /rust-toolchain-input/vllm \
+    && git fetch -v --prune -- origin ${VLLM_BRANCH} \
+    && git sparse-checkout init --no-cone \
+    && git sparse-checkout set \
+        rust-toolchain.toml \
+        tools/install_protoc.sh \
+    && git checkout FETCH_HEAD
+
+FROM rust_toolchain_input_${REMOTE_VLLM} AS rust-toolchain-input
+
+# -----------------------
+# Rust build inputs
+#
+# REMOTE_VLLM=0: use only the local files that can affect Rust artifacts.
+# REMOTE_VLLM=1: fetch only those same files from the requested remote ref.
+FROM scratch AS rust_input_0
+COPY requirements/build/rust.txt /rust-input/vllm/requirements/build/rust.txt
+COPY rust/Cargo.lock rust/Cargo.toml /rust-input/vllm/rust/
+COPY rust/proto /rust-input/vllm/rust/proto
+COPY rust/src /rust-input/vllm/rust/src
+COPY rust-toolchain.toml /rust-input/vllm/rust-toolchain.toml
+COPY tools/build_rust.py /rust-input/vllm/tools/build_rust.py
+COPY build_rust.sh /rust-input/vllm/build_rust.sh
+
+FROM base AS rust_input_1
+ARG VLLM_REPO
+ARG VLLM_BRANCH
+RUN git clone --no-checkout --filter=blob:none ${VLLM_REPO} /rust-input/vllm \
+    && cd /rust-input/vllm \
+    && git fetch -v --prune -- origin ${VLLM_BRANCH} \
+    && git sparse-checkout init --no-cone \
+    && git sparse-checkout set \
+        requirements/build/rust.txt \
+        rust/Cargo.lock \
+        rust/Cargo.toml \
+        rust/proto/** \
+        rust/src/** \
+        rust-toolchain.toml \
+        tools/build_rust.py \
+        build_rust.sh \
+    && git checkout FETCH_HEAD
+
+FROM rust_input_${REMOTE_VLLM} AS rust-input
+
+# -----------------------
+# Rust/protoc toolchain
+FROM base AS rust-toolchain
+
+ENV CARGO_HOME=/root/.cargo
+ENV RUSTUP_HOME=/root/.rustup
+ENV PATH=${CARGO_HOME}/bin:${PATH}

 # protoc is used by tonic-build/prost-build.
 RUN apt-get update -q -y && apt-get install -q -y --no-install-recommends \
        ca-certificates curl unzip \
    && rm -rf /var/lib/apt/lists/*

-COPY tools/install_protoc.sh /tmp/install_protoc.sh
+COPY --from=rust-toolchain-input /rust-toolchain-input/vllm/tools/install_protoc.sh /tmp/install_protoc.sh
 RUN /tmp/install_protoc.sh && rm /tmp/install_protoc.sh

+COPY --from=rust-toolchain-input /rust-toolchain-input/vllm/rust-toolchain.toml /tmp/rust-toolchain.toml
+RUN TOOLCHAIN="$(grep '^channel' /tmp/rust-toolchain.toml | sed 's/.*= *"\(.*\)"/\1/')" \
+    && if ! command -v rustup >/dev/null 2>&1; then \
+        curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs \
+            | sh -s -- -y --default-toolchain none; \
+    fi \
+    && rustup toolchain install "${TOOLCHAIN}" \
+    && rustup default "${TOOLCHAIN}" \
+    && rustc --version \
+    && cargo --version
+
+# -----------------------
+# Rust build stage
+# Builds the `vllm-rs` frontend in a dedicated stage so the wheel build stages
+# don't need the rust toolchain or protoc.
+FROM rust-toolchain AS rust-build
+ARG COMMON_WORKDIR
+ARG USE_SCCACHE
+
 # Cap cargo parallelism to avoid exhausting the AMD CI host's open-file limit
 # (rustc spawns enough concurrent processes to hit RLIMIT_NOFILE otherwise).
 ENV CARGO_BUILD_JOBS=4
 ENV CARGO_NET_RETRY=10
 ENV RUSTUP_MAX_RETRIES=10

+WORKDIR ${COMMON_WORKDIR}/vllm
+
+COPY --from=rust-input /rust-input/vllm/requirements/build/rust.txt requirements/build/rust.txt
+
 # BuildKit can run this stage in parallel with ROCm native builds. Keep Rust on
 # a separate local sccache daemon while sharing the same remote cache backend.
 ENV SCCACHE_SERVER_PORT=4227

 RUN --mount=type=cache,id=vllm-rocm-uv,target=/root/.cache/uv \
-    cd ${COMMON_WORKDIR}/vllm \
-    && uv pip install --system -r requirements/build/rust.txt
+    uv pip install --system -r requirements/build/rust.txt
+
+COPY --from=rust-input /rust-input/vllm/rust rust
+COPY --from=rust-input /rust-input/vllm/rust-toolchain.toml rust-toolchain.toml
+COPY --from=rust-input /rust-input/vllm/tools/build_rust.py tools/build_rust.py
+COPY --from=rust-input /rust-input/vllm/build_rust.sh build_rust.sh

 # Build the release binary. Cargo's registry/git caches can be written by
 # concurrent BuildKit jobs on shared workers, so lock those cache mounts while
@@ -159,8 +243,7 @@ RUN --mount=type=cache,id=vllm-rocm-uv,target=/root/.cache/uv \
 # can outlive source updates across BuildKit cache reuse.
 RUN --mount=type=cache,id=vllm-rocm-cargo-registry,target=/root/.cargo/registry,sharing=locked \
    --mount=type=cache,id=vllm-rocm-cargo-git,target=/root/.cargo/git,sharing=locked \
-    cd ${COMMON_WORKDIR}/vllm \
-    && if [ "$USE_SCCACHE" = "1" ]; then \
+    if [ "$USE_SCCACHE" = "1" ]; then \
        export RUSTC_WRAPPER=sccache \
        && sccache --show-stats; \
    fi \
@@ -185,8 +268,8 @@ COPY requirements/common.txt requirements/common.txt
 RUN --mount=type=cache,id=vllm-rocm-uv,target=/root/.cache/uv \
    uv pip install --system -r requirements/rocm.txt

-# pyproject.toml is bind-mounted in the RUN step so metadata-only changes do
-# not invalidate the expensive native build layer.
+# pyproject.toml is bind-mounted so it is not persisted in the csrc-build image,
+# but it still participates in the BuildKit cache key for this RUN step.
 COPY setup.py CMakeLists.txt ./
 COPY tools/build_rust.py tools/build_rust.py
 COPY cmake cmake/
@@ -556,6 +639,17 @@ COPY --from=build_vllm_wheel_release ${COMMON_WORKDIR}/vllm/vllm/v1 /vllm_v1
 FROM mori_base AS ci_base
 ARG COMMON_WORKDIR

+ENV CARGO_HOME=/root/.cargo
+ENV RUSTUP_HOME=/root/.rustup
+ENV PATH=${CARGO_HOME}/bin:${PATH}
+
+# Bake the Rust/protoc toolchain into ci_base so Rust build/test steps do not
+# need to fetch it on each per-commit image build.
+COPY --from=rust-toolchain /root/.cargo /root/.cargo
+COPY --from=rust-toolchain /root/.rustup /root/.rustup
+COPY --from=rust-toolchain /usr/local/bin/protoc /usr/local/bin/protoc
+COPY --from=rust-toolchain /usr/local/include/google /usr/local/include/google
+
 # Update rdma-core to support latest rocshmem.
 ARG DEEPEP_NIC
 RUN if [ "${DEEPEP_NIC}" = "cx7" ] || [ "${DEEPEP_NIC}" = "io" ]; then \
@@ -37,37 +37,6 @@ RUN microdnf install -y \
    python${PYTHON_VERSION}-devel python${PYTHON_VERSION}-pip python${PYTHON_VERSION}-wheel  && \
    python${PYTHON_VERSION} -m venv $VIRTUAL_ENV && pip install --no-cache -U pip wheel uv && microdnf clean all

-FROM python-install AS pyarrow
-
-# Build Apache Arrow
-WORKDIR /tmp
-RUN --mount=type=cache,target=/root/.cache/uv \
-    git clone https://github.com/apache/arrow.git  -b maint-19.0.1 && \
-    cd arrow/cpp && \
-    mkdir release && cd release && \
-    cmake -DCMAKE_BUILD_TYPE=Release \
-          -DCMAKE_INSTALL_PREFIX=/usr/local \
-          -DARROW_PYTHON=ON \
-          -DARROW_PARQUET=ON \
-          -DARROW_ORC=ON \
-          -DARROW_FILESYSTEM=ON \
-          -DARROW_WITH_LZ4=ON \
-          -DARROW_WITH_ZSTD=ON \
-          -DARROW_WITH_SNAPPY=ON \
-          -DARROW_JSON=ON \
-          -DARROW_CSV=ON \
-          -DARROW_DATASET=ON \
-          -DPROTOBUF_PROTOC_EXECUTABLE=/usr/bin/protoc \
-          -DARROW_DEPENDENCY_SOURCE=BUNDLED \
-          .. && \
-    make -j$(nproc) && \
-    make install && \
-    cd ../../python && \
-    export PYARROW_PARALLEL=4 && \
-    export ARROW_BUILD_TYPE=release && \
-    uv pip install -r requirements-build.txt && \
-    python setup.py build_ext --build-type=$ARROW_BUILD_TYPE --bundle-arrow-cpp bdist_wheel
-
 FROM python-install AS rust
 ENV CARGO_HOME=/root/.cargo
 ENV RUSTUP_HOME=/root/.rustup
@@ -121,60 +90,18 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 FROM python-install AS numba-builder

 ARG MAX_JOBS
-ARG NUMBA_VERSION=0.61.2
+ARG NUMBA_VERSION=0.65.0

 WORKDIR /tmp

 # Clone all required dependencies
 RUN --mount=type=cache,target=/root/.cache/uv \
    microdnf install ninja-build gcc gcc-c++ -y && \
-    git clone --recursive https://github.com/llvm/llvm-project.git -b llvmorg-15.0.7  && \
-    git clone --recursive https://github.com/numba/llvmlite.git -b v0.44.0 && \
+    git clone --recursive https://github.com/numba/llvmlite.git -b v0.47.0 && \
    git clone --recursive https://github.com/numba/numba.git -b ${NUMBA_VERSION} && \
-    cd llvm-project && mkdir build && cd  build && \
+    cd llvmlite && \
    uv pip install 'cmake<4' 'setuptools<70' numpy && \
-    export PREFIX=/usr/local && CMAKE_ARGS="${CMAKE_ARGS} -DLLVM_ENABLE_PROJECTS=lld;libunwind;compiler-rt" \
-    CFLAGS="$(echo $CFLAGS | sed 's/-fno-plt //g')" \
-    CXXFLAGS="$(echo $CXXFLAGS | sed 's/-fno-plt //g')" \
-    CMAKE_ARGS="${CMAKE_ARGS} -DFFI_INCLUDE_DIR=$PREFIX/include" \
-    CMAKE_ARGS="${CMAKE_ARGS} -DFFI_LIBRARY_DIR=$PREFIX/lib" \
-    cmake -DCMAKE_INSTALL_PREFIX="${PREFIX}"               \
-        -DCMAKE_BUILD_TYPE=Release                       \
-        -DCMAKE_LIBRARY_PATH="${PREFIX}"                 \
-        -DLLVM_ENABLE_LIBEDIT=OFF                        \
-        -DLLVM_ENABLE_LIBXML2=OFF                        \
-        -DLLVM_ENABLE_RTTI=ON                            \
-        -DLLVM_ENABLE_TERMINFO=OFF                       \
-        -DLLVM_INCLUDE_BENCHMARKS=OFF                    \
-        -DLLVM_INCLUDE_DOCS=OFF                          \
-        -DLLVM_INCLUDE_EXAMPLES=OFF                      \
-        -DLLVM_INCLUDE_GO_TESTS=OFF                      \
-        -DLLVM_INCLUDE_TESTS=OFF                         \
-        -DLLVM_INCLUDE_UTILS=ON                          \
-        -DLLVM_INSTALL_UTILS=ON                          \
-        -DLLVM_UTILS_INSTALL_DIR=libexec/llvm            \
-        -DLLVM_BUILD_LLVM_DYLIB=OFF                      \
-        -DLLVM_LINK_LLVM_DYLIB=OFF                       \
-        -DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD=WebAssembly \
-        -DLLVM_ENABLE_FFI=ON                             \
-        -DLLVM_ENABLE_Z3_SOLVER=OFF                      \
-        -DLLVM_OPTIMIZED_TABLEGEN=ON                     \
-        -DCMAKE_POLICY_DEFAULT_CMP0111=NEW               \
-        -DCOMPILER_RT_BUILD_BUILTINS=ON                  \
-        -DCOMPILER_RT_BUILTINS_HIDE_SYMBOLS=OFF          \
-        -DCOMPILER_RT_BUILD_LIBFUZZER=OFF                \
-        -DCOMPILER_RT_BUILD_CRT=OFF                      \
-        -DCOMPILER_RT_BUILD_MEMPROF=OFF                  \
-        -DCOMPILER_RT_BUILD_PROFILE=OFF                  \
-        -DCOMPILER_RT_BUILD_SANITIZERS=OFF               \
-        -DCOMPILER_RT_BUILD_XRAY=OFF                     \
-        -DCOMPILER_RT_BUILD_GWP_ASAN=OFF                 \
-        -DCOMPILER_RT_BUILD_ORC=OFF                      \
-        -DCOMPILER_RT_INCLUDE_TESTS=OFF                  \
-        ${CMAKE_ARGS} -GNinja ../llvm                    \
-    && ninja install  . && \
-    #  build llvmlite
-    cd ../../llvmlite && python setup.py bdist_wheel && \
+    python setup.py bdist_wheel && \
    cd ../numba && \
    if ! grep '#include "dynamic_annotations.h"' numba/_dispatcher.cpp; then \
       sed -i '/#include "internal\/pycore_atomic.h"/i\#include "dynamic_annotations.h"' numba/_dispatcher.cpp; \
@@ -192,21 +119,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
    cd opencv-python && \
    python -m build --wheel --installer=uv --outdir /tmp/opencv-python/dist

-## Todo(r3hankhan123): Remove guidance-builder stage once vLLM upgrades to new version of llguidance that fixes s390x issues. See https://github.com/guidance-ai/llguidance/issues/330
-FROM python-install AS guidance-builder
-WORKDIR /tmp
-ENV CARGO_HOME=/root/.cargo
-ENV RUSTUP_HOME=/root/.rustup
-ENV PATH="$CARGO_HOME/bin:$RUSTUP_HOME/bin:$PATH"
-RUN --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=bind,from=rust,source=/root/.cargo,target=/root/.cargo,rw \
-    --mount=type=bind,from=rust,source=/root/.rustup,target=/root/.rustup,rw \
-    git clone https://github.com/guidance-ai/llguidance.git && \
-    cd llguidance && \
-    git checkout s390x-fix-v2  && \
-    uv pip install maturin && \
-    python -m maturin build --release --out dist --compatibility linux
-
 # # Final build stage
 FROM python-install AS vllm-cpu
 ARG PYTHON_VERSION
@@ -235,20 +147,16 @@ RUN --mount=type=bind,from=numa-build,src=/tmp/numactl-2.0.19,target=/numactl \
 RUN --mount=type=cache,target=/root/.cache/uv \
    --mount=type=bind,from=rust,source=/root/.cargo,target=/root/.cargo,rw \
    --mount=type=bind,from=rust,source=/root/.rustup,target=/root/.rustup,rw \
-    --mount=type=bind,from=pyarrow,source=/tmp/arrow/python/dist,target=/tmp/arrow-wheels \
    --mount=type=bind,from=torch-vision,source=/tmp/vision/dist,target=/tmp/vision-wheels/ \
    --mount=type=bind,from=hf-xet-builder,source=/tmp/hf-xet/dist,target=/tmp/hf-xet-wheels/ \
    --mount=type=bind,from=numba-builder,source=/tmp/llvmlite/dist,target=/tmp/llvmlite-wheels/ \
    --mount=type=bind,from=numba-builder,source=/tmp/numba/dist,target=/tmp/numba-wheels/ \
    --mount=type=bind,from=opencv-builder,source=/tmp/opencv-python/dist,target=/tmp/opencv-wheels/ \
-    --mount=type=bind,from=guidance-builder,source=/tmp/llguidance/dist,target=/tmp/guidance-wheels/ \
-     ARROW_WHL_FILE=$(ls /tmp/arrow-wheels/*.whl) && \
     VISION_WHL_FILE=$(ls /tmp/vision-wheels/*.whl) && \
     HF_XET_WHL_FILE=$(ls /tmp/hf-xet-wheels/*.whl) && \
     LLVM_WHL_FILE=$(ls /tmp/llvmlite-wheels/*.whl) && \
     NUMBA_WHL_FILE=$(ls /tmp/numba-wheels/*.whl) && \
     OPENCV_WHL_FILE=$(ls /tmp/opencv-wheels/*.whl) && \
-     GUIDANCE_WHL_FILE=$(ls /tmp/guidance-wheels/*.whl) && \
     uv pip install -v \
        $ARROW_WHL_FILE \
        $VISION_WHL_FILE \
@@ -256,7 +164,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
        $LLVM_WHL_FILE \
        $NUMBA_WHL_FILE \
        $OPENCV_WHL_FILE \
-        $GUIDANCE_WHL_FILE \
        --torch-backend cpu \
        --index-strategy unsafe-best-match \
        -r requirements/build/cpu.txt \
@@ -141,6 +141,10 @@ variable "ROCM_CSRC_CACHE_TO_MODE" {
  default = "max"
 }

+variable "ROCM_RUST_CACHE_TO_MODE" {
+  default = "max"
+}
+
 variable "ROCM_FINAL_CACHE_TO_MODE" {
  default = "min"
 }
@@ -164,6 +168,13 @@ function "get_cache_from_rocm" {
    VLLM_MERGE_BASE_COMMIT != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:csrc-rocm-${VLLM_MERGE_BASE_COMMIT}" : "",
    ROCM_CACHE_BRANCH_TAG != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:csrc-rocm-branch-${ROCM_CACHE_BRANCH_TAG}" : "",
    ROCM_CACHE_UPSTREAM_BRANCH_TAG != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:csrc-rocm-branch-${ROCM_CACHE_UPSTREAM_BRANCH_TAG}" : "",
+    # Import the source-scoped Rust frontend cache so non-Rust changes do not
+    # force a fresh cargo release build.
+    BUILDKITE_COMMIT != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-${BUILDKITE_COMMIT}" : "",
+    PARENT_COMMIT != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-${PARENT_COMMIT}" : "",
+    VLLM_MERGE_BASE_COMMIT != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-${VLLM_MERGE_BASE_COMMIT}" : "",
+    ROCM_CACHE_BRANCH_TAG != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-branch-${ROCM_CACHE_BRANCH_TAG}" : "",
+    ROCM_CACHE_UPSTREAM_BRANCH_TAG != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-branch-${ROCM_CACHE_UPSTREAM_BRANCH_TAG}" : "",
    # Branch-scoped full image cache - fallback when parent-commit cache is evicted
    ROCM_CACHE_BRANCH_TAG != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rocm-branch-${ROCM_CACHE_BRANCH_TAG}" : "",
    ROCM_CACHE_UPSTREAM_BRANCH_TAG != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rocm-branch-${ROCM_CACHE_UPSTREAM_BRANCH_TAG}" : "",
@@ -204,6 +215,27 @@ function "get_cache_to_rocm_csrc" {
  ])
 }

+function "get_cache_from_rocm_rust" {
+  params = []
+  result = compact([
+    BUILDKITE_COMMIT != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-${BUILDKITE_COMMIT}" : "",
+    PARENT_COMMIT != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-${PARENT_COMMIT}" : "",
+    VLLM_MERGE_BASE_COMMIT != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-${VLLM_MERGE_BASE_COMMIT}" : "",
+    ROCM_CACHE_BRANCH_TAG != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-branch-${ROCM_CACHE_BRANCH_TAG}" : "",
+    ROCM_CACHE_UPSTREAM_BRANCH_TAG != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-branch-${ROCM_CACHE_UPSTREAM_BRANCH_TAG}" : "",
+  ])
+}
+
+function "get_cache_to_rocm_rust" {
+  params = []
+  result = compact([
+    # Export exact-commit and branch-scoped Rust caches. A content-addressed
+    # cache ref is appended by ci-bake-rocm.sh when that wrapper is used.
+    BUILDKITE_COMMIT != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-${BUILDKITE_COMMIT},mode=${ROCM_RUST_CACHE_TO_MODE}" : "",
+    ROCM_CACHE_BRANCH_TAG != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rust-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${ROCM_RUST_CACHE_TO_MODE}" : "",
+  ])
+}
+
 # Cache functions for upstream dependency stages (RIXL/UCX, ROCShmem, DeepEP).
 # These stages are pinned to specific upstream commit hashes, so cache keys use
 # those hashes rather than the Buildkite commit. This means the cache persists
@@ -276,6 +308,17 @@ target "csrc-rocm-ci" {
  output     = ["type=cacheonly"]
 }

+# Cache-only target for the Rust frontend build stage. Final-image cache
+# exports use mode=min and do not reliably persist intermediate cargo layers,
+# so Rust gets its own source-scoped cache target.
+target "rust-rocm-ci" {
+  inherits   = ["_common-rocm", "_ci-rocm"]
+  target     = "rust-build"
+  cache-from = get_cache_from_rocm_rust()
+  cache-to   = get_cache_to_rocm_rust()
+  output     = ["type=cacheonly"]
+}
+
 # Keep wheel export on the same CI graph as the test image build so the
 # shared build_vllm/export_vllm stages resolve identically within one bake
 # invocation. Without this, export-wheel-rocm uses the plain local target
@@ -292,13 +335,13 @@ target "export-wheel-rocm" {
 # Artifact-only vLLM build. GPU test jobs consume this artifact on top of
 # ci_base, avoiding a per-commit multi-GB image push/pull.
 group "test-rocm-ci-with-artifacts" {
-  targets = ["csrc-rocm-ci", "export-wheel-rocm"]
+  targets = ["rust-rocm-ci", "csrc-rocm-ci", "export-wheel-rocm"]
 }

 # Full test image + wheel export. Kept for fallback/debugging when a pushed
 # per-commit image is useful.
 group "test-rocm-ci-with-wheel" {
-  targets = ["csrc-rocm-ci", "test-rocm-ci", "export-wheel-rocm"]
+  targets = ["rust-rocm-ci", "csrc-rocm-ci", "test-rocm-ci", "export-wheel-rocm"]
 }

 # Image tags for the ci_base build. ci-bake-rocm.sh rewrites CI_BASE_IMAGE_TAG
@@ -232,7 +232,7 @@ MLA decode backends are selected using the standard
 | `ROCM_AITER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %1 | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
 | `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 1, 64 | Any | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
 | `ROCM_AITER_TRITON_MLA` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `TOKENSPEED_MLA` | fp16, bf16 | `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | 10.x |
+| `TOKENSPEED_MLA` | fp16, bf16 | `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
 | `TRITON_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
 | `XPU_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | Any | 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | Any |

@@ -358,6 +358,7 @@ th {
 | `DeepseekForCausalLM` | DeepSeek | `deepseek-ai/deepseek-llm-67b-base`, `deepseek-ai/deepseek-llm-7b-chat`, etc. | ✅︎ | ✅︎ |
 | `DeepseekV2ForCausalLM` | DeepSeek-V2 | `deepseek-ai/DeepSeek-V2`, `deepseek-ai/DeepSeek-V2-Chat`, etc. | ✅︎ | ✅︎ |
 | `DeepseekV3ForCausalLM` | DeepSeek-V3 | `deepseek-ai/DeepSeek-V3`, `deepseek-ai/DeepSeek-R1`, `deepseek-ai/DeepSeek-V3.1`, etc. | ✅︎ | ✅︎ |
+| `DeepseekV32ForCausalLM` | DeepSeek-V3.2 | `deepseek-ai/DeepSeek-V3.2`, etc. | ✅︎ | ✅︎ |
 | `DeepseekV4ForCausalLM` | DeepSeek-V4 | `deepseek-ai/DeepSeek-V4-Flash`, `deepseek-ai/DeepSeek-V4-Pro`, etc. | | ✅︎ |
 | `DotsOCRForCausalLM` | dots_ocr | `rednote-hilab/dots.ocr` | ✅︎ | ✅︎ |
 | `Ernie4_5ForCausalLM` | Ernie4.5 | `baidu/ERNIE-4.5-0.3B-PT`, etc. | ✅︎ | ✅︎ |
@@ -21,7 +21,7 @@ pillow  # Required for image processing
 prometheus-fastapi-instrumentator >= 8.0.0 # v8 unblocks starlette >= 1.0
 tiktoken >= 0.6.0  # Required for DBRX tokenizer
 lm-format-enforcer == 0.11.3
-llguidance >= 1.7.0, < 1.8.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64" or platform_machine == "ppc64le"
+llguidance >= 1.7.0, < 1.8.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64" or platform_machine == "ppc64le" or platform_machine == "s390x"
 outlines_core == 0.2.14
 # required for outlines backend disk cache
 diskcache == 5.6.3
@@ -26,7 +26,7 @@ nvidia-cutlass-dsl[cu13]==4.5.2
 quack-kernels>=0.3.3

 # Tokenspeed_MLA for faster mla with spec decode
-tokenspeed-mla==0.1.2; platform_system == "Linux"
+tokenspeed-mla==0.1.8; platform_system == "Linux"

 # Humming kernels for quantization gemm
 humming-kernels[cu13]==0.1.10
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 syntax = "proto3";
 package vllm;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::time::Duration;

 use anyhow::{Context, Result, bail};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::sync::Arc;

 use tracing::info;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::HashMap;
 use std::sync::Arc;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use thiserror::Error;
 use thiserror_ext::{AsReport as _, Macro};

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::ops::Deref;
 use std::sync::Arc;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Minimal chat facade above [`vllm_text`].
 //!
 //! This crate keeps the northbound boundary intentionally small:
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Chat-layer multimodal media preparation.
 //!
 //! This module owns the multimodal path for chat requests: it extracts media
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Prompt placeholder expansion shared across modalities.

 use std::collections::{HashMap, VecDeque};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Image-modality preparation: batch preprocessing and per-item feature
 //! build.

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::HashMap;

 use half::{bf16, f16};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Video-modality preparation: per-clip preprocessing, config resolution,
 //! and per-item feature build.

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Default output processing pipeline.

 mod structural_tag;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Applies xgrammar structural-tag constraints for strict tool calling.

 use thiserror_ext::AsReport;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Adapts decoded text updates into parsed assistant deltas.
 //!
 //! This stage sits between low-level token decoding and final block assembly.
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Native Harmony output processing for `gpt_oss`.
 //!
 //! Unlike the default text-first pipeline, this processor consumes
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::sync::Arc;

 use futures::executor::block_on;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::pin::Pin;
 use std::sync::Arc;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Adapts parsed assistant updates into structured chat events.
 //!
 //! This module remains the final assembly stage in `vllm-chat`. Token-to-text
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 pub mod reasoning;
 pub mod tool;
 pub mod unified;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Reasoning parser registration and selection boundary for `vllm-chat`.

 use std::sync::{Arc, LazyLock};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::sync::Arc;

 use vllm_tokenizer::test_utils::TestTokenizer;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Tool parser registration and selection boundary for `vllm-chat`.

 use std::sync::{Arc, LazyLock};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use vllm_parser::tool::{Result, ToolParserOutput};

 use super::{ToolParser, ToolParserFactory, names};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Unified parser registration and selection boundary for `vllm-chat`.

 use std::sync::LazyLock;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! DeepSeek V3.2 prompt renderer.

 use std::collections::{HashMap, HashSet};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 mod encoding;

 use vllm_text::Prompt;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::path::PathBuf;

 use expect_test::{ExpectFile, expect, expect_file};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! DeepSeek V4 prompt renderer.
 //!
 //! Original Python implementation:
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 mod encoding;

 use vllm_text::Prompt;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::path::PathBuf;

 use expect_test::{ExpectFile, expect, expect_file};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Shared Harmony encoding helper for the GPT-OSS renderer and output parser.

 use std::sync::LazyLock;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Native Harmony chat renderer for `gpt_oss`.

 pub(crate) mod encoding;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::path::PathBuf;

 use expect_test::{ExpectFile, expect, expect_file};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use thiserror::Error as ThisError;

 #[derive(Debug, ThisError)]
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::{HashSet, VecDeque};
 use std::fmt;
 use std::str::FromStr;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::HashMap;

 use serde::Serialize;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Chat template support for tokenizers using Jinja2 templates.
 //!
 //! This module is inlined from SMG's tokenizer crate with local adaptations:
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use minijinja::value::{Kwargs, ViaDeserialize};
 use minijinja::{Error as MinijinjaError, ErrorKind, Value};
 use serde::Deserialize;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::sync::Arc;

 use indexmap::IndexMap;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::HashMap;
 use std::sync::Arc;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::fmt;
 use std::str::FromStr;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::fs;
 use std::path::Path;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::HashMap;

 use llm_multimodal::ImageDetail;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::pin::Pin;
 use std::sync::Arc;
 use std::task::{Context, Poll};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::BTreeSet;
 use std::fmt;
 use std::sync::Arc;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! Roundtrip tests for the real chat-template and output-processor pairing.
 //!
 //! The invariant under test is that a structured assistant message rendered as history can be
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 //! CLI argument definitions for the `vllm-rs` binary.
 //!
 //! Python vLLM references:
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use expect_test::expect;
 use vllm_engine_core_client::TransportMode;
 use vllm_server::{Config, HttpListenerMode, ParserSelection, RendererSelection};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 #![allow(clippy::doc_lazy_continuation)]

 use std::fmt::Display;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::{env, fmt, process};

 use time::UtcOffset;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 mod cli;
 mod logging;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::time::Duration;

 use anyhow::{Context, Result, bail};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::time::Duration;

 use anyhow::{Context, Result, bail};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::sync::Arc;
 use std::time::Duration;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::{BTreeMap, BTreeSet};
 use std::sync::Arc;
 use std::time::{SystemTime, UNIX_EPOCH};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::{BTreeMap, BTreeSet, HashMap};
 use std::sync::atomic::{AtomicU64, Ordering};

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::ops::Deref;
 use std::pin::Pin;
 use std::task::{Context, Poll};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::time::Duration;

 use bytes::Bytes;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::sync::Arc;

 use serde_tuple::{Deserialize_tuple, Serialize_tuple};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::sync::Arc;

 use parking_lot::Mutex;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::sync::Arc;

 use serde_tuple::{Deserialize_tuple, Serialize_tuple};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 mod bootstrap;
 mod external;
 mod handle;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::sync::Arc;
 use std::time::Duration;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 mod client;
 mod coordinator;
 mod error;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::BTreeMap;
 use std::collections::BTreeSet;
 use std::time::{SystemTime, UNIX_EPOCH};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::path::Path;
 use std::time::Duration;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use serde::{Deserialize, Serialize};

 /// Effective model dtype reported by the engine after config resolution.
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::BTreeMap;

 use serde::{Deserialize, Serialize};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 mod array;
 #[cfg(test)]
 mod tests;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::io::Cursor;

 use byteorder::{BigEndian, LittleEndian, NativeEndian, ReadBytesExt};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::BTreeSet;

 use bytes::Bytes;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use serde_tuple::{Deserialize_tuple, Serialize_tuple};

 use crate::protocol::tensor::WireNdArray;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use serde_tuple::{Deserialize_tuple, Serialize_tuple};

 use crate::protocol::OpaqueValue;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::any::type_name;
 use std::io::Cursor;

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::BTreeMap;

 use serde::{Deserialize, Serialize};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::BTreeSet;

 use enum_as_inner::EnumAsInner;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::{BTreeMap, HashMap};

 use bytes::Bytes;
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::{BTreeSet, HashMap};

 use serde::{Deserialize, Serialize};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use std::collections::BTreeMap;

 use serde::{Deserialize, Serialize};
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 use enum_as_inner::EnumAsInner;
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 use serde_json::Value;
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
yewentao256	3a981fbbdb	fused padding to kernel Signed-off-by: yewentao256 <zhyanwentao@126.com>	2026-07-13 22:13:11 +00:00
Pavani Majety GitHub Jingyi Yang	7fc97042c3	Add DCP + Eagle support for Tokenspeed MLA backends (#48180 ) Signed-off-by: Pavani Majety <pmajety@nvidia.com> Signed-off-by: Jingyi Yang <girasoleyang@gmail.com> Co-authored-by: Jingyi Yang <girasoleyang@gmail.com>	2026-07-13 11:46:02 -07:00
Micah WilliamsonandGitHub	18c4067a54	[ROCm][CI] Unblock `AMD: Language Models Test (Extended Pooling)` (#48513 ) Signed-off-by: Micah Williamson <micah.williamson@amd.com>	2026-07-13 18:38:10 +00:00
akii96 GitHub Ben Browning	550218b136	[Bugfix][Frontend] Flush engine reasoning parser at engine-reasoning → tool streaming boundary (#47606 ) Signed-off-by: Aakif Nawaz <aakif.nawaz@amd.com> Co-authored-by: Ben Browning <56071+bbrowning@users.noreply.github.com>	2026-07-13 14:06:10 -04:00
Gavin MorrisandGitHub	5c342876a6	[Doc] Add DeepseekV32ForCausalLM to supported_models.md (#48293 ) Signed-off-by: Gavin Morris <gmorriscs@gmail.com>	2026-07-13 17:43:59 +00:00
stefankoncarevic GitHub Harry Mellor	9427c45386	[ROCm][CI] Transformers: pass only one of input_ids/inputs_embeds (#48258 ) Signed-off-by: Stefan Koncarevic <stefan.koncarevic@amd.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>	2026-07-13 17:28:50 +00:00
omerpaz95 GitHub Or Ozeri	43c8cbf79b	[EC Connector] CPU Offloading EC Connector (#47423 ) Signed-off-by: omerpaz95 <omerpaz95@gmail.com> Signed-off-by: Or Ozeri <oro@il.ibm.com> Co-authored-by: Or Ozeri <oro@il.ibm.com>	2026-07-13 20:09:41 +03:00
Taneem Ibrahim GitHubmergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	62286308c9	[Misc] Improve Matryoshka pooling dimensions validation (#48057 ) Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>	2026-07-13 12:57:36 -04:00
Nick HillandGitHub	26587f9519	[BugFix][ModelRunner V2] Fix stale attn metadata in speculator prefill cudagraph capture (#48261 ) Signed-off-by: Nick Hill <nickhill123@gmail.com>	2026-07-13 09:39:15 -07:00
xiangdong GitHub Kunshang Ji	93e3bc8f30	[XPU][CI]Adjust timeout_in_minutes in Intel GPU CI (#48418 ) Signed-off-by: zengxian <xiangdong.zeng@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>	2026-07-13 23:11:16 +08:00
Yan MaandGitHub	c2c9f7c5e2	remove force channels_last in Idefics3MultiModalProcessor (#48467 ) Signed-off-by: Yan Ma <yan.ma@intel.com>	2026-07-13 14:18:57 +00:00
Omer Ullman ArgovandGitHub	1be6e937b2	lower memory required for capturing cudagraphs for large cudagraph sizes (#48483 ) Signed-off-by: Omer Ullman Argov <118735753+omera-nv@users.noreply.github.com>	2026-07-13 10:14:25 -04:00
Wentao YeandGitHub	b3cfca996c	[Mypy Fix] Split mypy work (#48490 ) Signed-off-by: yewentao256 <zhyanwentao@126.com>	2026-07-13 12:42:42 +00:00
Bugen ZhaoandGitHub	487dfb3418	[CI] Add SPDX license header to Rust/Protobuf sources (#48472 ) Signed-off-by: Bugen Zhao <i@bugenzhao.com>	2026-07-13 10:22:47 +01:00
Karthik Kothuri GitHub Claude wang.yuqi	107a03ba63	[Core] Support fp32 lm_head for generation models via head_dtype (RFC #48305 §3.6) (#48390 ) Signed-off-by: Karthik Kothuri <karthikkothuri2009@gmail.com> Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io>	2026-07-13 16:43:34 +08:00
Nicolò Lucchesi GitHub Claude	56a357ed33	[Bugfix][KV Cache] Don't route uniform-page-size MLA+SWA models into DeepseekV4 packing (#48256 ) Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai> Co-authored-by: Claude <noreply@anthropic.com>	2026-07-13 08:16:24 +00:00
Nicolò Lucchesi GitHub Claude Opus 4.8 Andreas Karatzas	bea70c7cfc	[Attention] Make sliding-window support an explicit backend capability (#48011 ) Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>	2026-07-13 01:07:56 -07:00
Mohammad Miadh AngkadandGitHub	75fe92a316	[Distributed][Perf] Enable FlashInfer MNNVL allreduce RMS quant fusion (#48064 ) Signed-off-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>	2026-07-13 15:02:59 +08:00
Andreas Karatzas GitHub Matthew Wong	b7b58d1eba	[ROCm][CI] Cache Rust builds by source inputs (#46527 ) Signed-off-by: Andreas Karatzas <akaratza@amd.com> Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com> Co-authored-by: Matthew Wong <Matthew.Wong2@amd.com>	2026-07-13 01:14:08 -05:00
Canlin GuoandGitHub	36484e464a	[BugFix] Restore full tokens for Qwen MTP When MoE SP (#48429 ) Signed-off-by: gcanlin <canlinguosdu@gmail.com>	2026-07-13 13:29:41 +08:00
Rehan Khan GitHubLi, Jiang <jiang1.li@intel.com>	9e57de7197	[CPU] Create Proper Numa topology for s390x (#40714 ) Signed-off-by: Rehan Khan <Rehan.Khan7@ibm.com> Co-authored-by: Li, Jiang <jiang1.li@intel.com>	2026-07-13 12:58:43 +08:00
Yejing LaiandGitHub	8c5dafcd09	[Bugfix][UT]Fix EagleMiniCPMForCausalLM meet TypeError (#48452 ) Signed-off-by: Lai, Yejing <yejing.lai@intel.com>	2026-07-13 04:37:23 +00:00
guybd GitHub CursorLi, Jiang <jiang1.li@intel.com>	05fa8183a6	[CPU][Spec Decode] Support DFlash speculative decoding for GDN models on CPU (#46090 ) Signed-off-by: guybd <guy.boudoukh@intel.com> Signed-off-by: Guy Boudoukh <guy.boudoukh@intel.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Li, Jiang <jiang1.li@intel.com>	2026-07-13 04:16:18 +00:00
Rohan Potdar GitHub Claude Opus 4.8	d973cce3ca	Re-disable CUDA graph memory profiling on ROCm (#48440 ) Signed-off-by: Rohan Potdar <rohan.potdar@amd.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-07-13 03:59:20 +00:00
Fangzhou Ai GitHub Cursor	775c1589ea	[Bugfix][ROCm] Keep TP all_gather on base-class collective (#48446 ) Signed-off-by: fai <fangzhouai@gmail.com> Co-authored-by: Cursor <cursoragent@cursor.com>	2026-07-13 03:53:53 +00:00