kimi-k3

Co-authored-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Ziming Huang <zelda.huanghuang@gmail.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Isotr0py <mozf@inferact.ai>
Co-authored-by: aoshen02 <aoshen02@users.noreply.github.com>
Co-authored-by: Woosuk Kwon <woosuk@inferact.ai>
Co-authored-by: Jee Jee Li <jeejeelee@inferact.ai>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com>
Co-authored-by: Summer Yang <girasoleyang@gmail.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
Co-authored-by: khluu <khluu000@gmail.com>
Co-authored-by: Bowen Wang <abmfy@icloud.com>
Co-authored-by: gnovack <novackgm@gmail.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: xiaozhoupy <peiyuanzhou1994@gmail.com>
Co-authored-by: Roy Wang <yasong.wang@inferact.ai>
Co-authored-by: Jeff (Junze) Ma <93145857+majunze2001@users.noreply.github.com>
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>

.buildkite/hardware_tests/amd.yaml

@@ -18,6 +18,8 @@ steps:
       TERM: "xterm-256color"
     retry:
       automatic:
+        - exit_status: 1  # Transient Docker/BuildKit failure
+          limit: 1
         - exit_status: -1  # Agent was lost
           limit: 1
         - exit_status: -10  # Agent was lost
@@ -46,6 +48,8 @@ steps:
       VLLM_BRANCH: "$BUILDKITE_COMMIT"
     retry:
       automatic:
+        - exit_status: 1  # Transient Docker/BuildKit failure
+          limit: 1
         - exit_status: -1  # Agent was lost
           limit: 1
         - exit_status: -10  # Agent was lost
@@ -72,6 +76,8 @@ steps:
       VLLM_BRANCH: "$BUILDKITE_COMMIT"
     retry:
       automatic:
+        - exit_status: 1  # Transient Docker/BuildKit failure
+          limit: 1
         - exit_status: -1  # Agent was lost
           limit: 1
         - exit_status: -10  # Agent was lost

.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
@@ -141,9 +145,22 @@ steps:
   commands:
     - |
       bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
-      pytest -x -v -s tests/models/multimodal/generation --ignore=tests/models/multimodal/generation/test_pixtral.py -m cpu_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
+      pytest -x -v -s tests/models/multimodal/generation --ignore=tests/models/multimodal/generation/test_pixtral.py --ignore=tests/models/multimodal/generation/test_qwen2_5_vl.py -m cpu_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
   parallelism: 4
 
+- label: CPU-Qwen2.5-VL Multimodal Tests
+  depends_on: []
+  device: intel_cpu
+  no_plugin: true
+  source_file_dependencies:
+  # - vllm/
+  - vllm/model_executor/layers/rotary_embedding
+  - tests/models/multimodal/generation/
+  commands:
+    - |
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 40m "
+      VLLM_CI_ENV=0 pytest -x -v -s tests/models/multimodal/generation/test_qwen2_5_vl.py"
+
 - label: "Arm CPU Test"
   depends_on: []
   soft_fail: false

.buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml

@@ -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:

.buildkite/image_build/image_build.sh

@@ -177,6 +177,18 @@ BRANCH=$4
 IMAGE_TAG=$5
 IMAGE_TAG_LATEST=${6:-} # only used for main branch, optional
 
+# When TORCH_NIGHTLY=1, build the base CI image against PyTorch nightly so the
+# entire existing pipeline runs on nightly torch (CUDA/GPU lane only). Delegate
+# to the dedicated nightly build (PYTORCH_NIGHTLY=1, CUDA 13.0) and tag it at the
+# normal IMAGE_TAG that every test step already pulls -- no separate image tag,
+# no duplicate "vLLM Against PyTorch Nightly" pipeline section.
+if [[ "${TORCH_NIGHTLY:-0}" == "1" ]]; then
+    echo "--- :warning: TORCH_NIGHTLY=1 -- building base image on PyTorch nightly"
+    SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+    exec "${SCRIPT_DIR}/image_build_torch_nightly.sh" \
+        "${REGISTRY}" "${REPO}" "${BUILDKITE_COMMIT}" "${BRANCH}" "${IMAGE_TAG}"
+fi
+
 # build config
 TARGET="test-ci"
 VLLM_BAKE_FILE_PATH="${VLLM_BAKE_FILE_PATH:-docker/docker-bake.hcl}"

.buildkite/image_build/image_build_arm64.sh

@@ -19,13 +19,14 @@ if docker manifest inspect "$IMAGE" >/dev/null 2>&1; then
   echo "Image found"
 else
   echo "Image not found, proceeding with build..."
-  # build for arm64 GPU targets: Grace/GH200 (sm_90) and DGX Spark/GB10
+  # build for arm64 GPU targets: Grace/GH200 (sm_90),
+  # Blackwell/Thor (sm_100/sm_103/sm_110), and DGX Spark/GB10
   # (sm_121, family-covered by 12.0 under CUDA 13)
   docker build --file docker/Dockerfile \
     --platform linux/arm64 \
     --build-arg max_jobs=16 \
     --build-arg nvcc_threads=4 \
-    --build-arg torch_cuda_arch_list="9.0 12.0" \
+    --build-arg torch_cuda_arch_list="9.0 10.0 11.0 12.0" \
     --build-arg USE_SCCACHE=1 \
     --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
     --tag "$IMAGE" \

.buildkite/intel_jobs/basic_correctness.yaml

@@ -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

.buildkite/intel_jobs/lora_intel.yaml

@@ -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

.buildkite/intel_jobs/misc_intel.yaml

@@ -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
@@ -72,9 +72,7 @@ steps:
       pytest -v -s v1/test_oracle.py &&
       pytest -v -s v1/test_request.py &&
       pytest -v -s v1/test_outputs.py &&
-      pytest -v -s v1/sample/test_topk_topp_sampler.py &&
-      pytest -v -s v1/sample/test_logprobs.py &&
-      pytest -v -s v1/sample/test_logprobs_e2e.py'
+      pytest -v -s v1/sample'
 
 - label: Basic Models Tests (Initialization)
   timeout_in_minutes: 60
@@ -154,7 +152,7 @@ steps:
 
 - label: Regression
   key: regression
-  timeout_in_minutes: 30
+  timeout_in_minutes: 50
   device: intel_gpu
   agent_tags:
     label: production
@@ -188,7 +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:
@@ -224,7 +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

.buildkite/intel_jobs/models_distributed_intel.yaml

@@ -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

.buildkite/intel_jobs/models_multimodal_intel.yaml

@@ -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

.buildkite/intel_jobs/test-intel.yaml

@@ -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
@@ -99,12 +99,13 @@ steps:
         pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py &&
         pytest -v -s v1/structured_output &&
         pytest -v -s v1/test_serial_utils.py &&
+        pytest -v -s v1/e2e/general/test_correctness_sliding_window.py --deselect="tests/v1/e2e/general/test_correctness_sliding_window.py::test_sliding_window_retrieval[True-1-5-google/gemma-3-1b-it]" &&
         pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py --ignore=v1/spec_decode/test_speculators_correctness.py &&
         pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py --ignore=v1/kv_connector/unit/test_hf3fs_client.py --ignore=v1/kv_connector/unit/test_hf3fs_connector.py --ignore=v1/kv_connector/unit/test_hf3fs_metadata_server.py --ignore=v1/kv_connector/unit/test_offloading_connector.py'
   - label: "XPU server test"
     depends_on:
       - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 45
     device: intel_gpu
     agent_tags:
       label: production

.buildkite/lm-eval-harness/configs/Meta-Llama-4-Maverick-17B-128E-Instruct-FP8.yaml

@@ -1,6 +1,7 @@
 # For hf script, without -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-mmlupro-vllm-baseline.sh -m meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 -l 250 -t 8 -f 5
 model_name: "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
+rocm_safetensors_load_strategy: lazy
 required_gpu_arch:
   - gfx942
   - gfx950

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -72,6 +72,11 @@ def launch_lm_eval(eval_config, tp_size):
     if moe_backend is not None:
         model_args += f"moe_backend={moe_backend},"
 
+    if current_platform.is_rocm():
+        rocm_load_strategy = eval_config.get("rocm_safetensors_load_strategy")
+        if rocm_load_strategy is not None:
+            model_args += f"safetensors_load_strategy={rocm_load_strategy},"
+
     env_vars = eval_config.get("env_vars", None)
     with scoped_env_vars(env_vars):
         results = lm_eval.simple_evaluate(

.buildkite/scripts/annotate-build-artifact.sh

@@ -3,7 +3,8 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 #
 # Append a build artifact line to the Buildkite annotation.
-# Usage: annotate-build-artifact.sh <label> <value>
+# Usage: annotate-build-artifact.sh <label> <value> <context>
 set -e
 echo "- **${1}**: \`${2}\`" | \
-  buildkite-agent annotate --append --style 'info' --context 'release-artifacts'
+  buildkite-agent annotate --append --style 'info' \
+    --context "${3:?context is required}"

.buildkite/scripts/build-macos-wheel.sh

@@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Build the macOS arm64 CPU wheel natively on a macOS agent (the `macmini`
+# queue) into artifacts/dist/ for upload-nightly-wheels.sh.
+
+set -euo pipefail
+
+# The Rust frontend build needs protoc.
+if ! command -v protoc >/dev/null 2>&1; then
+  brew install protobuf
+fi
+
+# upload-nightly-wheels.sh expects exactly one wheel.
+rm -rf artifacts/dist
+mkdir -p artifacts/dist
+
+export VLLM_TARGET_DEVICE=cpu
+export VLLM_REQUIRE_RUST_FRONTEND=1
+export MACOSX_DEPLOYMENT_TARGET=11.0
+# uv's CPython is universal2; force an arm64-only build and tag so the wheel
+# isn't mislabelled universal2 and installed on Intel Macs where import fails.
+export ARCHFLAGS="-arch arm64"
+export _PYTHON_HOST_PLATFORM="macosx-11.0-arm64"
+export CMAKE_BUILD_PARALLEL_LEVEL="${CMAKE_BUILD_PARALLEL_LEVEL:-4}"
+
+uv venv --python 3.12
+uv pip install -r requirements/build/cpu.txt --index-strategy unsafe-best-match
+uv build --wheel --no-build-isolation -o artifacts/dist
+
+ls -l artifacts/dist/*.whl

.buildkite/scripts/check-ray-compatibility.sh

@@ -29,7 +29,11 @@ if python3 -c "import torch; assert torch.version.hip" 2>/dev/null; then
         TORCH_INDEX_URL=""
     fi
 else
-    TORCH_INDEX_URL="https://download.pytorch.org/whl/cu130"
+    if [ "${TORCH_NIGHTLY:-0}" = "1" ]; then
+        TORCH_INDEX_URL="https://download.pytorch.org/whl/nightly/cu130"
+    else
+        TORCH_INDEX_URL="https://download.pytorch.org/whl/cu130"
+    fi
 fi
 echo ">>> Using PyTorch index: ${TORCH_INDEX_URL:-PyPI default}"
 

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

@@ -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_nixl build_rocshmem build_deepep mori_base ci_base"
 DEFAULT_CI_BASE_METADATA_VERSION="1"
 IMAGE_EXISTED_BEFORE_BUILD=0
 
@@ -285,7 +285,7 @@ get_content_arg_names() {
     fi | awk 'NF && !seen[$0]++'
 }
 
-compute_ci_base_content_hash() {
+compute_ci_base_content_hash_once() {
     local -a content_paths=()
     local -a content_args=()
     local dockerfile="${CI_BASE_DOCKERFILE:-}"
@@ -301,7 +301,8 @@ compute_ci_base_content_hash() {
         if [[ -n "${dockerfile}" ]]; then
             printf 'dockerfile:%s\n' "${dockerfile}"
             printf 'resolved-build-args:\n'
-            hash_dockerfile_arg_values "${dockerfile}" "${content_args[@]}"
+            hash_dockerfile_arg_values "${dockerfile}" "${content_args[@]}" \
+                || return 1
             if [[ -n "${stages}" ]]; then
                 printf 'dockerfile-stages:%s\n' "${stages}"
                 if [[ -f "${dockerfile}" ]]; then
@@ -314,6 +315,53 @@ compute_ci_base_content_hash() {
     } | sha256sum | cut -d' ' -f1
 }
 
+compute_ci_base_content_hash() {
+    local attempts="${CI_BASE_HASH_ATTEMPTS:-3}"
+    local delay_secs="${CI_BASE_HASH_RETRY_DELAY:-5}"
+    local attempt=0
+    local hash=""
+    local failed=0
+    local -a hashes=()
+
+    if [[ ! "${attempts}" =~ ^[1-9][0-9]*$ ]]; then
+        echo "Invalid CI_BASE_HASH_ATTEMPTS: ${attempts}" >&2
+        return 1
+    fi
+    if [[ ! "${delay_secs}" =~ ^[0-9]+$ ]]; then
+        echo "Invalid CI_BASE_HASH_RETRY_DELAY: ${delay_secs}" >&2
+        return 1
+    fi
+
+    for ((attempt = 1; attempt <= attempts; attempt++)); do
+        if ! hash=$(compute_ci_base_content_hash_once); then
+            echo "ci_base content hash calculation ${attempt}/${attempts} failed" >&2
+            failed=1
+        else
+            hashes+=("${hash}")
+            echo "ci_base content hash calculation ${attempt}/${attempts}: ${hash}" >&2
+        fi
+
+        if ((attempt < attempts)); then
+            sleep "${delay_secs}"
+        fi
+    done
+
+    if ((failed)) || ((${#hashes[@]} != attempts)); then
+        echo "Could not calculate a reliable ci_base content hash" >&2
+        return 1
+    fi
+
+    for hash in "${hashes[@]:1}"; do
+        if [[ "${hash}" != "${hashes[0]}" ]]; then
+            echo "ci_base content hash changed between calculations" >&2
+            printf '  observed: %s\n' "${hashes[@]}" >&2
+            return 1
+        fi
+    done
+
+    printf '%s\n' "${hashes[0]}"
+}
+
 extract_dockerfile_arg_default() {
     local dockerfile="$1"
     local arg_name="$2"
@@ -366,7 +414,11 @@ hash_dockerfile_arg_values() {
         printf 'arg:%s=%s\n' "${arg_name}" "${arg_value:-<empty>}"
         if [[ "${arg_name}" == "BASE_IMAGE" && -n "${arg_value}" ]]; then
             digest=$(resolve_image_digest "${arg_value}")
-            printf 'arg:%s.digest=%s\n' "${arg_name}" "${digest:-unknown}"
+            if [[ -z "${digest}" ]]; then
+                echo "Failed to resolve digest for BASE_IMAGE=${arg_value}" >&2
+                return 1
+            fi
+            printf 'arg:%s.digest=%s\n' "${arg_name}" "${digest}"
         fi
     done
 }
@@ -764,7 +816,7 @@ configure_ci_base_image_refs() {
         fi
         set_buildkite_metadata "rocm-ci-base-image" "${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
@@ -1107,8 +1159,8 @@ ci_base_metadata_pairs() {
     metadata_pair "vllm.rocm.nic_backend" "$(resolve_dockerfile_arg_value "${dockerfile}" "NIC_BACKEND")"
     metadata_pair "vllm.rocm.ainic_version" "$(resolve_dockerfile_arg_value "${dockerfile}" "AINIC_VERSION")"
     metadata_pair "vllm.rocm.ubuntu_codename" "$(resolve_dockerfile_arg_value "${dockerfile}" "UBUNTU_CODENAME")"
-    metadata_pair "vllm.rocm.rixl_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "RIXL_REPO")"
-    metadata_pair "vllm.rocm.rixl_commit" "${RIXL_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "RIXL_BRANCH")}"
+    metadata_pair "vllm.rocm.nixl_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "NIXL_REPO")"
+    metadata_pair "vllm.rocm.nixl_commit" "${NIXL_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "NIXL_BRANCH")}"
     metadata_pair "vllm.rocm.ucx_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "UCX_REPO")"
     metadata_pair "vllm.rocm.ucx_commit" "${UCX_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "UCX_BRANCH")}"
     metadata_pair "vllm.rocm.rocshmem_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "ROCSHMEM_REPO")"
@@ -1117,7 +1169,7 @@ ci_base_metadata_pairs() {
     metadata_pair "vllm.rocm.deepep_commit" "${DEEPEP_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_BRANCH")}"
     metadata_pair "vllm.rocm.deepep_nic" "$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_NIC")"
     metadata_pair "vllm.rocm.deepep_rocm_arch" "$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_ROCM_ARCH")"
-    metadata_pair "vllm.rocm.rixl_cache_key" "${RIXL_CACHE_KEY:-}"
+    metadata_pair "vllm.rocm.nixl_cache_key" "${NIXL_CACHE_KEY:-}"
     metadata_pair "vllm.rocm.rocshmem_cache_key" "${ROCSHMEM_CACHE_KEY:-}"
     metadata_pair "vllm.rocm.deepep_cache_key" "${DEEPEP_CACHE_KEY:-}"
 
@@ -1211,12 +1263,24 @@ 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 +1324,56 @@ 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="$1"
     local value=""
@@ -1317,6 +1431,7 @@ write_rocm_build_arg_override() {
                 "${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 +1480,133 @@ 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=${rocm_cache_to_mode},ignore-error=true"
         )
@@ -1422,7 +1624,7 @@ target "csrc-rocm-ci" {
   cache-from = concat(
     get_cache_from_rocm_csrc(),
 EOF
-        write_hcl_string_list "    " "${content_cache_from[@]}"
+        write_hcl_string_list "    " "${csrc_content_cache_from[@]}"
         cat <<EOF
   )
 EOF
@@ -1430,11 +1632,23 @@ 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 +1660,7 @@ 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
@@ -1472,7 +1686,7 @@ extract_dependency_pins() {
         return 0
     fi
 
-    for var in RIXL_BRANCH UCX_BRANCH ROCSHMEM_BRANCH DEEPEP_BRANCH; do
+    for var in NIXL_BRANCH UCX_BRANCH ROCSHMEM_BRANCH DEEPEP_BRANCH; do
         if [[ -n "${!var:-}" ]]; then
             echo "Using provided ${var}: ${!var}"
             continue
@@ -1492,30 +1706,30 @@ extract_dependency_pins() {
 compute_dependency_cache_keys() {
     local bake_dir=""
     local dockerfile_rocm=""
-    local rixl_branch=""
+    local nixl_branch=""
     local ucx_branch=""
     local rocshmem_branch=""
     local deepep_branch=""
-    local rixl_material=""
+    local nixl_material=""
     local rocshmem_material=""
     local deepep_material=""
 
     bake_dir=$(dirname "${VLLM_BAKE_FILE}")
     dockerfile_rocm="${bake_dir}/Dockerfile.rocm"
-    rixl_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "RIXL_BRANCH")
+    nixl_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "NIXL_BRANCH")
     ucx_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "UCX_BRANCH")
     rocshmem_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "ROCSHMEM_BRANCH")
     deepep_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "DEEPEP_BRANCH")
 
-    if [[ -n "${rixl_branch}" && -n "${ucx_branch}" ]]; then
-        rixl_material=$(compose_stage_cache_material "${dockerfile_rocm}" "base build_rixl")
-        RIXL_CACHE_KEY=$(
+    if [[ -n "${nixl_branch}" && -n "${ucx_branch}" ]]; then
+        nixl_material=$(compose_stage_cache_material "${dockerfile_rocm}" "base build_nixl")
+        NIXL_CACHE_KEY=$(
             compose_dependency_cache_key \
-                "${rixl_branch}-ucx-${ucx_branch}" \
-                "${rixl_material}"
+                "${nixl_branch}-ucx-${ucx_branch}" \
+                "${nixl_material}"
         )
-        export RIXL_CACHE_KEY
-        echo "RIXL dependency cache key: ${RIXL_CACHE_KEY}"
+        export NIXL_CACHE_KEY
+        echo "NIXL dependency cache key: ${NIXL_CACHE_KEY}"
     fi
 
     if [[ -n "${rocshmem_branch}" ]]; then
@@ -1566,11 +1780,11 @@ dependency_cache_ref_for_target() {
     local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
 
     case "${target}" in
-        rixl-rocm-ci)
-            if [[ -n "${RIXL_CACHE_KEY:-}" ]]; then
-                printf '%s\n' "${cache_repo}:rixl-rocm-${RIXL_CACHE_KEY}"
-            elif [[ -n "${RIXL_BRANCH:-}" ]]; then
-                printf '%s\n' "${cache_repo}:rixl-rocm-${RIXL_BRANCH}-ucx-${UCX_BRANCH:-}"
+        nixl-rocm-ci)
+            if [[ -n "${NIXL_CACHE_KEY:-}" ]]; then
+                printf '%s\n' "${cache_repo}:nixl-rocm-${NIXL_CACHE_KEY}"
+            elif [[ -n "${NIXL_BRANCH:-}" ]]; then
+                printf '%s\n' "${cache_repo}:nixl-rocm-${NIXL_BRANCH}-ucx-${UCX_BRANCH:-}"
             fi
             ;;
         rocshmem-rocm-ci)
@@ -1601,7 +1815,7 @@ add_dependency_cache_target() {
 
 resolve_ci_base_dependency_targets() {
     local mode="${ROCM_DEP_CACHE_EXPORT_MODE:-missing}"
-    local rixl_ref=""
+    local nixl_ref=""
     local rocshmem_ref=""
     local deepep_ref=""
 
@@ -1610,7 +1824,7 @@ resolve_ci_base_dependency_targets() {
     case "${mode}" in
         always)
             echo "ROCM_DEP_CACHE_EXPORT_MODE=always; exporting all dependency caches serially"
-            for target in rixl-rocm-ci rocshmem-rocm-ci deepep-rocm-ci; do
+            for target in nixl-rocm-ci rocshmem-rocm-ci deepep-rocm-ci; do
                 if [[ -n "$(dependency_cache_ref_for_target "${target}")" ]]; then
                     add_dependency_cache_target "${target}"
                 fi
@@ -1630,13 +1844,13 @@ resolve_ci_base_dependency_targets() {
             ;;
     esac
 
-    if [[ "${mode}" != "always" && -n "${RIXL_CACHE_KEY:-}" ]]; then
-        rixl_ref=$(dependency_cache_ref_for_target "rixl-rocm-ci")
-        if dependency_cache_ref_exists "${rixl_ref}"; then
-            echo "RIXL dependency cache exists: ${rixl_ref}"
+    if [[ "${mode}" != "always" && -n "${NIXL_CACHE_KEY:-}" ]]; then
+        nixl_ref=$(dependency_cache_ref_for_target "nixl-rocm-ci")
+        if dependency_cache_ref_exists "${nixl_ref}"; then
+            echo "NIXL dependency cache exists: ${nixl_ref}"
         else
-            echo "RIXL dependency cache missing; will seed: ${rixl_ref}"
-            add_dependency_cache_target "rixl-rocm-ci"
+            echo "NIXL dependency cache missing; will seed: ${nixl_ref}"
+            add_dependency_cache_target "nixl-rocm-ci"
         fi
     fi
 
@@ -1736,8 +1950,8 @@ confirm_remote_image_push() {
         fi
 
         if [[ -z "${remote_revision}" \
-              && ${IMAGE_EXISTED_BEFORE_BUILD} -eq 0 \
-              && image_tag_is_commit_scoped ]]; then
+              && ${IMAGE_EXISTED_BEFORE_BUILD} -eq 0 ]] \
+            && image_tag_is_commit_scoped; then
             echo "Remote image exists under a commit-scoped tag; accepting push despite missing revision label."
             return 0
         fi
@@ -1867,36 +2081,57 @@ upload_wheel_artifacts_if_present() {
     local wheel_dir="./wheel-export"
     local artifact_dir="artifacts/vllm-rocm-install"
     local archive_name="vllm-rocm-install.tar.gz"
+    local metadata_dir="${wheel_dir}/.vllm-ci-artifact"
+    local native_base_image=""
     local whl=""
     local whl_name=""
+    local -a wheels=()
 
     if ! should_upload_wheel_artifacts; then
         return 0
     fi
 
-    if [[ ! -d "${wheel_dir}" ]] || ! ls "${wheel_dir}"/*.whl >/dev/null 2>&1; then
-        echo "No ROCm wheel artifacts found in ${wheel_dir}"
-        return 0
+    if [[ -d "${wheel_dir}" ]]; then
+        mapfile -t wheels < <(find "${wheel_dir}" -maxdepth 1 -type f -name '*.whl' -print)
+    fi
+    if [[ ${#wheels[@]} -ne 1 ]]; then
+        echo "Expected exactly one ROCm wheel in ${wheel_dir}; found ${#wheels[@]}" >&2
+        return 1
+    fi
+    whl="${wheels[0]}"
+    whl_name=$(basename "${whl}")
+    native_base_image="${CI_BASE_IMAGE_TAG_COMMIT_REF:-${CI_BASE_IMAGE:-}}"
+    if [[ -z "${native_base_image}" ]]; then
+        echo "Native ROCm artifact requires a ci_base image reference" >&2
+        return 1
     fi
 
     echo "--- :package: Uploading ROCm vLLM install artifact"
-    mkdir -p "${artifact_dir}"
+    rm -rf "${artifact_dir}" "${metadata_dir}"
+    mkdir -p "${artifact_dir}" "${metadata_dir}"
+
+    printf '%s\n' "${BUILDKITE_COMMIT:-local}" > "${metadata_dir}/commit.txt"
+    printf '%s\n' "${native_base_image}" > "${metadata_dir}/native-base-image.txt"
+    printf '%s\n' "${CI_BASE_IMAGE:-}" > "${metadata_dir}/ci-base-image.txt"
+    printf '%s\n' "${IMAGE_TAG:-}" > "${metadata_dir}/fallback-image.txt"
+    printf '%s\n' "${whl_name}" > "${metadata_dir}/wheel-filename.txt"
 
     tar -C "${wheel_dir}" -czf "${artifact_dir}/${archive_name}" .
+    (
+        cd "${artifact_dir}"
+        sha256sum "${archive_name}" > "${archive_name}.sha256"
+    )
     echo "Created ${archive_name}: $(du -sh "${artifact_dir}/${archive_name}" | cut -f1)"
-    printf '%s\n' "${CI_BASE_IMAGE:-}" > "${artifact_dir}/ci-base-image.txt"
-    printf '%s\n' "${IMAGE_TAG:-}" > "${artifact_dir}/fallback-image.txt"
-
-    for whl in "${wheel_dir}"/*.whl; do
-        [[ -f "${whl}" ]] || continue
-        whl_name=$(basename "${whl}")
-        cp "${whl}" "${artifact_dir}/${whl_name}"
-        echo "Copied ${whl_name}: $(du -sh "${artifact_dir}/${whl_name}" | cut -f1)"
-    done
+    cp "${metadata_dir}"/*.txt "${artifact_dir}/"
+    cp "${whl}" "${artifact_dir}/${whl_name}"
+    echo "Copied ${whl_name}: $(du -sh "${artifact_dir}/${whl_name}" | cut -f1)"
 
     if command -v buildkite-agent >/dev/null 2>&1; then
-        buildkite-agent artifact upload "${artifact_dir}/*"
+        buildkite-agent artifact upload "${artifact_dir}/*" || return 1
         echo "ROCm vLLM install artifacts uploaded to ${artifact_dir}/"
+    elif [[ "${BUILDKITE:-false}" == "true" ]]; then
+        echo "buildkite-agent not found; cannot upload required ROCm artifacts" >&2
+        return 1
     else
         echo "Not in Buildkite, skipping artifact upload"
     fi
@@ -1920,6 +2155,7 @@ 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
@@ -1927,6 +2163,11 @@ main() {
         echo "BAKE_PRINT_ONLY=1 set; skipping build"
         return 0
     fi
+    if should_upload_wheel_artifacts; then
+        # wheel-export is an output directory, not a BuildKit cache. Starting
+        # clean prevents a failed/retried export from packaging a stale wheel.
+        rm -rf ./wheel-export
+    fi
     seed_dependency_caches_if_needed
     run_bake
     upload_wheel_artifacts_if_present

.buildkite/scripts/generate-and-upload-nightly-index.sh

@@ -45,8 +45,10 @@ $PYTHON .buildkite/scripts/generate-nightly-index.py --version "$SUBPATH" --curr
 echo "Uploading indices to $S3_COMMIT_PREFIX"
 aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "$S3_COMMIT_PREFIX"
 
-# copy to /nightly/ only if it is on the main branch and not a PR
-if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]]; then
+# copy to /nightly/ only when enabled for a main branch build that is not a PR
+if [[ "${UPDATE_NIGHTLY_INDEX:-1}" == "1" && \
+      "$BUILDKITE_BRANCH" == "main" && \
+      "$BUILDKITE_PULL_REQUEST" == "false" ]]; then
     echo "Uploading indices to overwrite /nightly/"
     aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/nightly/"
 fi
@@ -67,7 +69,7 @@ pure_version="${version%%+*}"
 echo "Pure version (without variant): $pure_version"
 
 # re-generate and copy to /<pure_version>/ only if it does not have "dev" in the version
-if [[ "$version" != *"dev"* ]]; then
+if [[ "${UPDATE_VERSION_INDEX:-1}" == "1" && "$version" != *"dev"* ]]; then
     echo "Re-generating indices for /$pure_version/"
     rm -rf "${INDICES_OUTPUT_DIR:?}"
     mkdir -p "$INDICES_OUTPUT_DIR"

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-# This script runs tests inside the corresponding ROCm docker container.
-# It handles both single-node and multi-node test configurations.
+# This script runs ROCm tests either directly in a native CI pod or inside the
+# corresponding Docker container. Multi-node tests continue to use Docker.
 #
 # Multi-node detection: Instead of matching on fragile group names, we detect
 # multi-node jobs structurally by looking for the bracket command syntax
@@ -34,10 +34,27 @@ set -o pipefail
 : "${CLICOLOR_FORCE:=1}"
 : "${PY_COLORS:=1}"
 : "${ROCM_DOCKER_TTY:=1}"
+: "${PYTHONFAULTHANDLER:=1}"
+: "${PYTEST_TIMEOUT:=2400}"
 if [[ " ${PYTEST_ADDOPTS:-} " != *" --color"* ]]; then
   PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--color=yes"
 fi
-export BUILDKIT_PROGRESS TERM FORCE_COLOR CLICOLOR_FORCE PY_COLORS PYTEST_ADDOPTS ROCM_DOCKER_TTY
+if [[ " ${PYTEST_ADDOPTS:-} " != *" --durations="* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--durations=25"
+fi
+if [[ " ${PYTEST_ADDOPTS:-} " != *" --durations-min="* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--durations-min=1.0"
+fi
+# Dump stacks after 25 minutes, then stop an individual test after 40 minutes.
+if [[ " ${PYTEST_ADDOPTS:-} " != *" faulthandler_timeout="* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }-o faulthandler_timeout=1500"
+fi
+if [[ " ${PYTEST_ADDOPTS:-} " != *" --timeout-method="* &&
+  " ${PYTEST_ADDOPTS:-} " != *" --timeout-method "* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--timeout-method=thread"
+fi
+export BUILDKIT_PROGRESS TERM FORCE_COLOR CLICOLOR_FORCE PY_COLORS PYTEST_ADDOPTS PYTEST_TIMEOUT ROCM_DOCKER_TTY
+export PYTHONFAULTHANDLER
 
 # Export Python path for commands that run directly on the host. Containerized
 # tests set this to /vllm-workspace below so spawned Python processes do not
@@ -53,6 +70,28 @@ report_docker_usage() {
   docker system df || true
 }
 
+clear_ci_orchestration_env() {
+  unset -v \
+    VLLM_TEST_GROUP_NAME \
+    VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE \
+    VLLM_CI_ARTIFACT_STEP \
+    VLLM_TEST_CACHE \
+    VLLM_CI_EXECUTION_MODE \
+    VLLM_CI_WORKSPACE \
+    VLLM_CI_REQUIRE_WORKSPACE_MOUNT \
+    VLLM_TEST_COMMANDS \
+    VLLM_CI_BRANCH \
+    VLLM_CI_BASE_IMAGE \
+    VLLM_CI_FALLBACK_IMAGE \
+    VLLM_CI_DOCKER_DISABLED \
+    VLLM_CI_ARTIFACT_GLOB \
+    VLLM_CI_ARTIFACT_CHECKSUM_GLOB \
+    VLLM_CI_EXPECTED_GPU_COUNT \
+    VLLM_CI_USE_ARTIFACTS \
+    VLLM_CI_RESULTS_ROOT \
+    VLLM_ALLOW_DEPRECATED_BEAM_SEARCH
+}
+
 cleanup_network() {
   local max_nodes=${NUM_NODES:-2}
   for node in $(seq 0 $((max_nodes - 1))); do
@@ -145,7 +184,11 @@ prepare_artifact_image() {
   fi
 
   cp "${wheel_dir}"/*.whl "${context_dir}/wheels/" || return 1
-  tar -C "${wheel_dir}" --exclude='*.whl' -cf - . \
+  tar -C "${wheel_dir}" \
+    --exclude='*.whl' \
+    --exclude='.vllm-ci-artifact' \
+    --exclude='./.vllm-ci-artifact' \
+    -cf - . \
     | tar -C "${workspace_dir}" -xf - || return 1
   cat > "${context_dir}/Dockerfile" <<'EOF'
 ARG BASE_IMAGE
@@ -168,6 +211,259 @@ EOF
   return 0
 }
 
+is_native_runtime() {
+  [[ "${AMD_CI_RUNTIME:-}" == "native" || "${NATIVE_CI:-}" == "true" ]]
+}
+
+validate_native_workspace() {
+  local workspace_dir="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
+  local workspace_real=""
+  local checkout_real=""
+  local workspace_mount=""
+
+  mkdir -p "${workspace_dir}" || return 1
+  workspace_real=$(readlink -m "${workspace_dir}") || return 1
+  if [[ -n "${BUILDKITE_BUILD_CHECKOUT_PATH:-}" ]]; then
+    checkout_real=$(readlink -m "${BUILDKITE_BUILD_CHECKOUT_PATH}") || return 1
+    if [[ "${checkout_real}" == "${workspace_real}" \
+      || "${checkout_real}" == "${workspace_real}/"* \
+      || "${workspace_real}" == "${checkout_real}/"* ]]; then
+      echo "Refusing to replace ${workspace_real}; it overlaps the Buildkite checkout ${checkout_real}" >&2
+      return 1
+    fi
+  fi
+  if [[ "${VLLM_CI_REQUIRE_WORKSPACE_MOUNT:-1}" == "1" ]]; then
+    if ! command -v findmnt >/dev/null 2>&1; then
+      echo "findmnt is required to verify the native workspace mount" >&2
+      return 1
+    fi
+    workspace_mount=$(findmnt -n -T "${workspace_real}" -o TARGET 2>/dev/null || true)
+    if [[ "$(readlink -m "${workspace_mount:-/}")" != "${workspace_real}" ]]; then
+      echo "Native CI requires a dedicated volume mounted at ${workspace_real}" >&2
+      return 1
+    fi
+  fi
+}
+
+prepare_native_workspace() {
+  if [[ "${VLLM_CI_USE_ARTIFACTS:-0}" != "1" ]]; then
+    echo "Native CI requires VLLM_CI_USE_ARTIFACTS=1"
+    return 1
+  fi
+  if ! command -v buildkite-agent >/dev/null 2>&1; then
+    echo "buildkite-agent not found; cannot download ROCm wheel artifact"
+    return 1
+  fi
+  validate_native_workspace || return 1
+
+  local artifact_glob="${VLLM_CI_ARTIFACT_GLOB:-artifacts/vllm-rocm-install/vllm-rocm-install.tar.gz}"
+  local artifact_checksum_glob="${VLLM_CI_ARTIFACT_CHECKSUM_GLOB:-${artifact_glob}.sha256}"
+  local artifact_step="${VLLM_CI_ARTIFACT_STEP:-image-build-amd}"
+  local archive=""
+  local checksum=""
+  local download_dir=""
+  local metadata_dir=""
+  local recorded_base=""
+  local recorded_commit=""
+  local recorded_wheel=""
+  local workspace_dir="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
+  local wheel_dir=""
+  local attempt=0
+  local attempt_dir=""
+  local -a archives=()
+  local -a checksums=()
+  local -a wheels=()
+
+  artifact_work_dir=$(mktemp -d -t vllm-rocm-artifact.XXXXXX) || return 1
+  wheel_dir="${artifact_work_dir}/wheels"
+  mkdir -p "${wheel_dir}" || return 1
+
+  echo "--- Downloading ROCm wheel artifact from ${artifact_step} (native in-pod)"
+  for attempt in 1 2 3; do
+    attempt_dir="${artifact_work_dir}/download-${attempt}"
+    rm -rf "${attempt_dir}" || return 1
+    mkdir -p "${attempt_dir}" || return 1
+    if buildkite-agent artifact download \
+      "${artifact_glob}" "${attempt_dir}" --step "${artifact_step}" \
+      && buildkite-agent artifact download \
+        "${artifact_checksum_glob}" "${attempt_dir}" --step "${artifact_step}"; then
+      download_dir="${attempt_dir}"
+      break
+    fi
+    echo "Artifact download attempt ${attempt}/3 failed"
+    if [[ "${attempt}" -lt 3 ]]; then
+      sleep $((attempt * 2))
+    fi
+  done
+  if [[ -z "${download_dir}" ]]; then
+    echo "Failed to download ${artifact_glob} and ${artifact_checksum_glob} from ${artifact_step}"
+    return 1
+  fi
+
+  mapfile -t archives < <(
+    find "${download_dir}" -name "vllm-rocm-install.tar.gz" -type f -print
+  )
+  mapfile -t checksums < <(
+    find "${download_dir}" -name "vllm-rocm-install.tar.gz.sha256" -type f -print
+  )
+  if [[ ${#archives[@]} -ne 1 || ${#checksums[@]} -ne 1 ]]; then
+    echo "Expected exactly one ROCm archive and checksum; found ${#archives[@]} archive(s) and ${#checksums[@]} checksum(s)" >&2
+    return 1
+  fi
+  archive="${archives[0]}"
+  checksum="${checksums[0]}"
+  if [[ "$(dirname "${archive}")" != "$(dirname "${checksum}")" ]]; then
+    echo "ROCm archive and checksum were downloaded to different directories" >&2
+    return 1
+  fi
+  (
+    cd "$(dirname "${archive}")"
+    sha256sum -c "$(basename "${checksum}")"
+  ) || return 1
+
+  tar --no-same-owner -xzf "${archive}" -C "${wheel_dir}" || return 1
+  mapfile -t wheels < <(
+    find "${wheel_dir}" -maxdepth 1 -type f -name '*.whl' -print
+  )
+  if [[ ${#wheels[@]} -ne 1 ]]; then
+    echo "ROCm artifact must contain exactly one top-level wheel; found ${#wheels[@]}" >&2
+    return 1
+  fi
+  metadata_dir="${wheel_dir}/.vllm-ci-artifact"
+  for metadata_file in commit.txt native-base-image.txt wheel-filename.txt; do
+    if [[ ! -s "${metadata_dir}/${metadata_file}" ]]; then
+      echo "ROCm artifact metadata is missing ${metadata_file}" >&2
+      return 1
+    fi
+  done
+  for metadata_file in ci-base-image.txt fallback-image.txt; do
+    if [[ ! -f "${metadata_dir}/${metadata_file}" ]]; then
+      echo "ROCm artifact metadata is missing ${metadata_file}" >&2
+      return 1
+    fi
+  done
+
+  recorded_commit=$(tr -d '\r\n' < "${metadata_dir}/commit.txt")
+  recorded_base=$(tr -d '\r\n' < "${metadata_dir}/native-base-image.txt")
+  recorded_wheel=$(tr -d '\r\n' < "${metadata_dir}/wheel-filename.txt")
+  if [[ -z "${BUILDKITE_COMMIT:-}" || "${recorded_commit}" != "${BUILDKITE_COMMIT}" ]]; then
+    echo "ROCm artifact commit ${recorded_commit} does not match ${BUILDKITE_COMMIT:-unset}" >&2
+    return 1
+  fi
+  if [[ -z "${VLLM_CI_BASE_IMAGE:-}" || "${recorded_base}" != "${VLLM_CI_BASE_IMAGE}" ]]; then
+    echo "ROCm artifact base ${recorded_base} does not match ${VLLM_CI_BASE_IMAGE:-unset}" >&2
+    return 1
+  fi
+  if [[ "${recorded_wheel}" != "$(basename "${wheels[0]}")" ]]; then
+    echo "ROCm artifact wheel manifest ${recorded_wheel} does not match $(basename "${wheels[0]}")" >&2
+    return 1
+  fi
+  for required_dir in tests .buildkite requirements; do
+    if [[ ! -d "${wheel_dir}/${required_dir}" ]]; then
+      echo "ROCm wheel artifact did not contain ${required_dir}/" >&2
+      return 1
+    fi
+  done
+
+  echo "--- Installing ROCm wheel into pod environment"
+  python3 -m pip install --no-deps --force-reinstall "${wheels[0]}" || return 1
+
+  echo "--- Preparing ${workspace_dir} from artifact"
+  find "${workspace_dir}" -mindepth 1 -maxdepth 1 -exec rm -rf -- {} + || return 1
+  tar -C "${wheel_dir}" \
+    --exclude='*.whl' \
+    --exclude='.vllm-ci-artifact' \
+    --exclude='./.vllm-ci-artifact' \
+    -cf - . | tar --no-same-owner -C "${workspace_dir}" -xf - || return 1
+  if [[ ! -d "${workspace_dir}/tests" ]]; then
+    echo "Failed to stage the native test workspace" >&2
+    return 1
+  fi
+
+  return 0
+}
+
+initialize_native_environment() {
+  local job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-local}}"
+  local job_id_suffix=""
+  local native_root=""
+  local hf_mount=""
+
+  if [[ "$(id -u)" -ne 0 ]]; then
+    echo "Native ROCm CI currently requires the ci_base container to run as root" >&2
+    return 1
+  fi
+
+  job_id="${job_id//[^A-Za-z0-9_.-]/_}"
+  job_id_suffix="${job_id##*-}"
+  job_id_suffix="${job_id_suffix:0:12}"
+  native_root="/tmp/vllm-native-${job_id}"
+  TMPDIR="/tmp/vllm-${job_id_suffix}/tmp"
+  VLLM_RPC_BASE_PATH="/tmp"
+  TORCHINDUCTOR_CACHE_DIR="${native_root}/cache/torchinductor"
+  TRITON_CACHE_DIR="${native_root}/cache/triton"
+  VLLM_CACHE_ROOT="${native_root}/cache/vllm"
+  XDG_CACHE_HOME="${native_root}/cache/xdg"
+  : "${HF_HOME:=/home/buildkite-agent/huggingface}"
+  : "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
+  : "${HF_HUB_ETAG_TIMEOUT:=60}"
+  export TMPDIR VLLM_RPC_BASE_PATH
+  export TORCHINDUCTOR_CACHE_DIR TRITON_CACHE_DIR VLLM_CACHE_ROOT XDG_CACHE_HOME
+  export HF_HOME HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
+  export PYTORCH_ROCM_ARCH=""
+
+  mkdir -p "${TMPDIR}" \
+    "${TORCHINDUCTOR_CACHE_DIR}" \
+    "${TRITON_CACHE_DIR}" \
+    "${VLLM_CACHE_ROOT}" \
+    "${XDG_CACHE_HOME}" \
+    "${HF_HOME}" || return 1
+
+  echo "Native compile caches: VLLM_CACHE_ROOT=${VLLM_CACHE_ROOT} TORCHINDUCTOR_CACHE_DIR=${TORCHINDUCTOR_CACHE_DIR}"
+
+  if [[ "${VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE:-0}" == "1" ]]; then
+    if ! command -v findmnt >/dev/null 2>&1; then
+      echo "findmnt is required to verify the native Hugging Face cache mount" >&2
+      return 1
+    fi
+    hf_mount=$(findmnt -n -T "${HF_HOME}" -o TARGET 2>/dev/null || true)
+    if [[ -z "${hf_mount}" || "${hf_mount}" == "/" ]]; then
+      echo "Native CI requires a persistent volume mounted at or above ${HF_HOME}" >&2
+      return 1
+    fi
+  fi
+}
+
+run_native_preflight() {
+  local expected_gpus="${VLLM_CI_EXPECTED_GPU_COUNT:-1}"
+
+  if [[ ! "${expected_gpus}" =~ ^[0-9]+$ ]]; then
+    echo "Invalid VLLM_CI_EXPECTED_GPU_COUNT=${expected_gpus}" >&2
+    return 1
+  fi
+
+  python3 -c "import encodings, importlib.metadata as im, importlib.util as iu; [im.version(d) for d in ('transformers', 'torch', 'ray', 'sympy', 'markupsafe', 'vllm')]; missing=[m for m in ('torch.utils.model_zoo', 'transformers.models.nomic_bert', 'ray.dag', 'sympy.physics', 'markupsafe._speedups') if iu.find_spec(m) is None]; assert not missing, missing" || return 1
+
+  if [[ "${expected_gpus}" == "0" ]]; then
+    echo "Native CPU-only AMD job: skipping ROCm device validation"
+    return 0
+  fi
+
+  echo "--- ROCm info"
+  rocminfo || return 1
+  VLLM_CI_EXPECTED_GPU_COUNT="${expected_gpus}" python3 - <<'PY'
+import os
+
+import torch
+
+expected = int(os.environ["VLLM_CI_EXPECTED_GPU_COUNT"])
+assert torch.version.hip, "PyTorch is not a ROCm build"
+assert torch.cuda.is_available(), "ROCm GPU is not available to PyTorch"
+actual = torch.cuda.device_count()
+assert actual == expected, f"Expected {expected} ROCm GPU(s), found {actual}"
+PY
+}
+
 is_multi_node() {
   local cmds="$1"
   # Primary signal: NUM_NODES environment variable set by the pipeline
@@ -350,7 +646,58 @@ re_quote_pytest_markers() {
 # Main
 ###############################################################################
 
-# --- GPU initialization ---
+if is_native_runtime; then
+  echo "--- Native in-pod ROCm CI (AMD_CI_RUNTIME=${AMD_CI_RUNTIME:-unset}, NATIVE_CI=${NATIVE_CI:-unset})"
+  artifact_work_dir=""
+
+  cleanup_native_workspace() {
+    if [[ -n "${artifact_work_dir}" ]]; then
+      rm -rf "${artifact_work_dir}"
+    fi
+  }
+  trap cleanup_native_workspace EXIT
+
+  if [[ -n "${VLLM_TEST_COMMANDS:-}" ]]; then
+    commands="${VLLM_TEST_COMMANDS}"
+    commands_source="env"
+  else
+    commands="$*"
+    commands_source="argv"
+    if [[ -z "$commands" ]]; then
+      echo "Error: No test commands provided for native CI." >&2
+      exit 1
+    fi
+  fi
+
+  if [[ "$commands_source" == "argv" ]]; then
+    commands=$(re_quote_pytest_markers "$commands")
+  fi
+
+  if is_multi_node "$commands"; then
+    echo "Native CI does not support multi-node jobs yet."
+    exit 1
+  fi
+
+  if ! initialize_native_environment; then
+    echo "Failed to initialize the native test environment"
+    exit 1
+  fi
+  if ! prepare_native_workspace; then
+    echo "Failed to prepare native test workspace"
+    exit 1
+  fi
+
+  export PYTHONPATH="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
+
+  echo "Native test commands: $commands"
+  run_native_preflight || exit 1
+  # Keep AMD CI orchestration variables out of vLLM's runtime environment.
+  clear_ci_orchestration_env
+  /bin/bash -o pipefail -c "${commands}"
+  handle_pytest_exit "$?"
+fi
+
+# --- GPU initialization for legacy Docker execution ---
 echo "--- ROCm info"
 rocminfo
 
@@ -452,25 +799,27 @@ fi
 
 echo "Final commands: $commands"
 
-# The ROCm test image often ships /vllm-workspace without .git (artifact tarball unpack).
-# tests/standalone_tests/python_only_compile.sh uses merge-base(HEAD, origin/main) for
-# wheels.vllm.ai; compute on the agent (full git checkout) and pass into the container.
-vllm_standalone_merge_base=""
-checkout="${BUILDKITE_BUILD_CHECKOUT_PATH:-}"
-if [[ -z "${checkout}" || ! -d "${checkout}" ]]; then
-  checkout="."
+standalone_merge_base_env=()
+if [[ "$commands" == *python_only_compile.sh* ]]; then
+  # The ROCm test image often ships /vllm-workspace without .git. Resolve the
+  # wheels.vllm.ai commit from the agent checkout for this test only.
+  vllm_standalone_merge_base=""
+  checkout="${BUILDKITE_BUILD_CHECKOUT_PATH:-}"
+  if [[ -z "${checkout}" || ! -d "${checkout}" ]]; then
+    checkout="."
+  fi
+  # Pass safe.directory per-command because Buildkite uses mixed user IDs.
+  if git -c "safe.directory=${checkout}" -C "${checkout}" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
+    vllm_standalone_merge_base="$(
+      git -c "safe.directory=${checkout}" -C "${checkout}" merge-base HEAD origin/main 2>/dev/null || true
+    )"
+  fi
+  if [[ -z "${vllm_standalone_merge_base}" ]]; then
+    vllm_standalone_merge_base="${BUILDKITE_COMMIT:-}"
+  fi
+  echo "INFO: passing CI_STANDALONE_MERGE_BASE into container: ${vllm_standalone_merge_base}"
+  standalone_merge_base_env=(-e "CI_STANDALONE_MERGE_BASE=${vllm_standalone_merge_base}")
 fi
-# Pass safe.directory per-command (-c) because buildkite runs will always fail
-# the next check on git 2.35.2+ due to mixed uses of root and buildkite-agent/uids.
-if git -c "safe.directory=${checkout}" -C "${checkout}" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
-  vllm_standalone_merge_base="$(
-    git -c "safe.directory=${checkout}" -C "${checkout}" merge-base HEAD origin/main 2>/dev/null || true
-  )"
-fi
-if [[ -z "${vllm_standalone_merge_base}" ]]; then
-  vllm_standalone_merge_base="${BUILDKITE_COMMIT:-}"
-fi
-echo "INFO: passing VLLM_STANDALONE_MERGE_BASE into container: ${vllm_standalone_merge_base}"
 
 MYPYTHONPATH="/vllm-workspace"
 
@@ -501,6 +850,7 @@ else
 fi
 
 # --- Route: multi-node vs single-node ---
+clear_ci_orchestration_env
 if is_multi_node "$commands"; then
   echo "--- Multi-node job detected"
   export DCKR_VER=$(docker --version | sed 's/Docker version \(.*\), build .*/\1/')
@@ -589,7 +939,9 @@ else
     -e FORCE_COLOR \
     -e CLICOLOR_FORCE \
     -e PY_COLORS \
+    -e PYTHONFAULTHANDLER \
     -e PYTEST_ADDOPTS \
+    -e PYTEST_TIMEOUT \
     -v "${HF_CACHE}:${HF_MOUNT}" \
     -e "HF_HOME=${HF_MOUNT}" \
     -e "PYTHONPATH=${MYPYTHONPATH}" \
@@ -599,7 +951,7 @@ else
     -e "VLLM_CACHE_ROOT=${CONTAINER_CACHE_ROOT}/vllm" \
     -e "XDG_CACHE_HOME=${CONTAINER_CACHE_ROOT}/xdg" \
     -e "PYTORCH_ROCM_ARCH=" \
-    -e "VLLM_STANDALONE_MERGE_BASE=${vllm_standalone_merge_base}" \
+    "${standalone_merge_base_env[@]}" \
     --name "${container_name}" \
     "${image_name}" \
     /bin/bash -c "${CONTAINER_PREFLIGHT} && ${commands}"

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

@@ -1,10 +1,11 @@
 #!/bin/bash
 set -euox pipefail
 
-export VLLM_CPU_KVCACHE_SPACE=1 
+export VLLM_CPU_KVCACHE_SPACE=1
 export VLLM_CPU_CI_ENV=1
-# Reduce sub-processes for acceleration
-export TORCH_COMPILE_DISABLE=1 
+# Skip torch.compile via vLLM's --enforce-eager flag (passed below) instead of
+# TORCH_COMPILE_DISABLE=1, which torch 2.12 no longer treats as a silent no-op
+# when callers specify fullgraph=True.
 export VLLM_ENABLE_V1_MULTIPROCESSING=0
 
 SDE_ARCHIVE="sde-external-10.7.0-2026-02-18-lin.tar.xz"
@@ -49,15 +50,15 @@ wait_for_pid_and_check_log() {
 }
 
 # Test Sky Lake (AVX512F)
-./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_0.log 2>&1 &
+./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_0.log 2>&1 &
 PID_TEST_0=$!
 
 # Test Cascade Lake (AVX512F + VNNI)
-./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_1.log 2>&1 &
+./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_1.log 2>&1 &
 PID_TEST_1=$!
 
 # Test Cooper Lake (AVX512F + VNNI + BF16)
-./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_2.log 2>&1 &
+./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_2.log 2>&1 &
 PID_TEST_2=$!
 
 wait_for_pid_and_check_log $PID_TEST_0 test_0.log

.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/scripts/hardware_ci/run-intel-ci-test.sh

@@ -35,6 +35,7 @@ case "${test_suite}" in
     pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py
     pytest -v -s v1/structured_output
     pytest -v -s v1/test_serial_utils.py
+    pytest -v -s v1/e2e/general/test_correctness_sliding_window.py --deselect="tests/v1/e2e/general/test_correctness_sliding_window.py::test_sliding_window_retrieval[True-1-5-google/gemma-3-1b-it]"
     pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py --ignore=v1/spec_decode/test_speculators_correctness.py
     pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py --ignore=v1/kv_connector/unit/test_hf3fs_client.py --ignore=v1/kv_connector/unit/test_hf3fs_connector.py --ignore=v1/kv_connector/unit/test_hf3fs_metadata_server.py --ignore=v1/kv_connector/unit/test_offloading_connector.py
     ;;

.buildkite/scripts/publish-release-images.sh

@@ -130,6 +130,22 @@ docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${ROCM_BASE_CACHE_KEY}-rocm
 docker push vllm/vllm-openai-rocm:latest-base
 docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
 
+# ---- XPU ----
+
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-xpu
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-xpu vllm/vllm-openai-xpu:latest-x86_64
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-xpu vllm/vllm-openai-xpu:v${RELEASE_VERSION}-x86_64
+docker push vllm/vllm-openai-xpu:latest-x86_64
+docker push vllm/vllm-openai-xpu:v${RELEASE_VERSION}-x86_64
+
+docker manifest rm vllm/vllm-openai-xpu:latest || true
+docker manifest rm vllm/vllm-openai-xpu:v${RELEASE_VERSION} || true
+docker manifest create vllm/vllm-openai-xpu:latest vllm/vllm-openai-xpu:latest-x86_64 --amend
+docker manifest create vllm/vllm-openai-xpu:v${RELEASE_VERSION} vllm/vllm-openai-xpu:v${RELEASE_VERSION}-x86_64 --amend
+docker manifest push vllm/vllm-openai-xpu:latest
+docker manifest push vllm/vllm-openai-xpu:v${RELEASE_VERSION}
+
 # ---- CPU ----
 # CPU images are behind separate block steps and may not have been built.
 # All-or-nothing: inspect both arches first, then either publish everything

.buildkite/scripts/rocm/build-test-image.sh

@@ -13,6 +13,18 @@ metadata_get() {
     fi
 }
 
+use_ci_base_if_present() {
+    local ci_base_image=""
+
+    ci_base_image="$(metadata_get rocm-ci-base-image)"
+    if [[ -z "${ci_base_image}" ]]; then
+        return 1
+    fi
+
+    export CI_BASE_IMAGE="${ci_base_image}"
+    echo "Using ROCm ci_base image selected by the preceding build step: ${CI_BASE_IMAGE}"
+}
+
 use_refreshed_base_if_present() {
     local base_refreshed=""
 
@@ -22,15 +34,12 @@ use_refreshed_base_if_present() {
     fi
 
     export BASE_IMAGE
-    export CI_BASE_IMAGE
     export IMAGE_TAG_LATEST
 
     BASE_IMAGE="$(metadata_get rocm-base-image)"
-    CI_BASE_IMAGE="$(metadata_get rocm-ci-base-image)"
     IMAGE_TAG_LATEST="$(metadata_get rocm-ci-image-descriptive)"
 
     echo "Using refreshed ROCm base image for test image: ${BASE_IMAGE}"
-    echo "Using refreshed ROCm ci_base image for test image: ${CI_BASE_IMAGE}"
     if [[ -n "${IMAGE_TAG_LATEST}" ]]; then
         echo "Also tagging full ROCm CI image as: ${IMAGE_TAG_LATEST}"
     fi
@@ -41,6 +50,8 @@ use_refreshed_base_if_present() {
 main() {
     local base_refreshed=0
 
+    use_ci_base_if_present || true
+
     if use_refreshed_base_if_present; then
         base_refreshed=1
     fi

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

@@ -21,16 +21,20 @@ export CARGO_HOME="${CARGO_HOME:-$HOME/.cargo}"
 export RUSTUP_HOME="${RUSTUP_HOME:-$HOME/.rustup}"
 export PATH="$CARGO_HOME/bin:$PATH"
 
+PROTOC_VERSION="${PROTOC_VERSION:-31.1}"
+CARGO_BINSTALL_VERSION="${CARGO_BINSTALL_VERSION:-1.20.1}"
+UV_VERSION="${UV_VERSION:-0.11.28}"
+PYO3_PYTHON_VERSION="${PYO3_PYTHON_VERSION:-3.12}"
+
+CARGO_SORT_VERSION_REQ="${CARGO_SORT_VERSION_REQ:-2}"
+CARGO_DENY_VERSION_REQ="${CARGO_DENY_VERSION_REQ:-0.20}"
+CARGO_NEXTEST_VERSION_REQ="${CARGO_NEXTEST_VERSION_REQ:-0.9}"
+
 log_section() {
   echo "--- $*"
 }
 
 install_protoc() {
-  if command -v protoc >/dev/null 2>&1; then
-    return
-  fi
-
-  local version="${PROTOC_VERSION:-31.1}"
   local arch
   case "$(uname -m)" in
     x86_64)
@@ -45,16 +49,17 @@ install_protoc() {
       ;;
   esac
 
-  local url="https://github.com/protocolbuffers/protobuf/releases/download/v${version}/protoc-${version}-linux-${arch}.zip"
+  local url="https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-${arch}.zip"
   local tmp_dir
   tmp_dir="$(mktemp -d)"
 
-  log_section "Installing protoc ${version}"
+  log_section "Installing protoc ${PROTOC_VERSION}"
   curl -L --proto '=https' --tlsv1.2 -sSf "$url" -o "$tmp_dir/protoc.zip"
   mkdir -p "$CARGO_HOME/bin"
   unzip -q "$tmp_dir/protoc.zip" bin/protoc 'include/*' -d "$CARGO_HOME"
   chmod +x "$CARGO_HOME/bin/protoc"
   rm -rf "$tmp_dir"
+  protoc --version
 }
 
 rust_toolchain() {
@@ -75,66 +80,48 @@ install_rust_toolchain() {
 }
 
 install_cargo_binstall() {
-  if command -v cargo-binstall >/dev/null 2>&1; then
-    return
-  fi
-
-  log_section "Installing cargo-binstall"
+  log_section "Installing cargo-binstall ${CARGO_BINSTALL_VERSION}"
   curl -L --proto '=https' --tlsv1.2 -sSf \
-    https://raw.githubusercontent.com/cargo-bins/cargo-binstall/main/install-from-binstall-release.sh \
-    | bash
+    "https://raw.githubusercontent.com/cargo-bins/cargo-binstall/v${CARGO_BINSTALL_VERSION}/install-from-binstall-release.sh" \
+    | env BINSTALL_VERSION="$CARGO_BINSTALL_VERSION" bash
+  cargo-binstall -V
 }
 
 install_cargo_sort() {
-  if command -v cargo-sort >/dev/null 2>&1; then
-    return
-  fi
-
-  log_section "Installing cargo-sort"
-  install_cargo_binstall
-  cargo binstall --no-confirm cargo-sort
+  log_section "Installing cargo-sort ${CARGO_SORT_VERSION_REQ}"
+  cargo binstall --no-confirm --force "cargo-sort@${CARGO_SORT_VERSION_REQ}"
 }
 
 install_cargo_deny() {
-  if command -v cargo-deny >/dev/null 2>&1; then
-    return
-  fi
-
-  log_section "Installing cargo-deny"
-  install_cargo_binstall
-  cargo binstall --no-confirm cargo-deny
+  log_section "Installing cargo-deny ${CARGO_DENY_VERSION_REQ}"
+  cargo binstall --no-confirm --force "cargo-deny@${CARGO_DENY_VERSION_REQ}"
 }
 
 install_cargo_nextest() {
-  if command -v cargo-nextest >/dev/null 2>&1; then
-    return
-  fi
-
-  log_section "Installing cargo-nextest"
-  install_cargo_binstall
-  cargo binstall --no-confirm --secure cargo-nextest
+  log_section "Installing cargo-nextest ${CARGO_NEXTEST_VERSION_REQ}"
+  cargo binstall \
+    --no-confirm \
+    --force \
+    --secure \
+    "cargo-nextest@${CARGO_NEXTEST_VERSION_REQ}"
 }
 
 install_uv() {
-  if command -v uv >/dev/null 2>&1; then
-    return
-  fi
-
-  log_section "Installing uv"
-  curl -LsSf --proto '=https' --tlsv1.2 https://astral.sh/uv/install.sh \
+  log_section "Installing uv ${UV_VERSION}"
+  curl -L --proto '=https' --tlsv1.2 -sSf \
+    "https://github.com/astral-sh/uv/releases/download/${UV_VERSION}/uv-installer.sh" \
     | env UV_INSTALL_DIR="$CARGO_HOME/bin" sh
+  uv --version
 }
 
 setup_pyo3_python() {
-  local python_version="${PYO3_PYTHON_VERSION:-3.12}"
-
-  log_section "Installing Python ${python_version} for PyO3 tests"
-  uv python install "$python_version"
+  log_section "Installing Python ${PYO3_PYTHON_VERSION} for PyO3 tests"
+  uv python install "$PYO3_PYTHON_VERSION"
   PYO3_PYTHON="$(uv python find \
     --managed-python \
     --no-project \
     --resolve-links \
-    "$python_version")"
+    "$PYO3_PYTHON_VERSION")"
   export PYO3_PYTHON
 
   local python_libdir
@@ -156,6 +143,7 @@ PY
 }
 
 run_style_clippy() {
+  install_cargo_binstall
   install_cargo_sort
   install_cargo_deny
 
@@ -186,6 +174,7 @@ run_style_clippy() {
 run_tests() {
   install_uv
   setup_pyo3_python
+  install_cargo_binstall
   install_cargo_nextest
 
   log_section "Running cargo nextest"

.buildkite/scripts/trigger-ci-build.sh

@@ -23,6 +23,7 @@ NC='\033[0m' # No Color
 # Default configuration
 PIPELINE="ci"
 DRY_RUN=true
+TORCH_NIGHTLY=false
 
 usage() {
     cat <<EOF
@@ -34,12 +35,14 @@ Sets RUN_ALL=1 and NIGHTLY=1 environment variables.
 SAFETY: Dry-run by default. Use --execute to actually trigger a build.
 
 Options:
-    --execute       Actually trigger the build (default: dry-run)
-    --pipeline      Buildkite pipeline slug (default: ${PIPELINE})
-    --commit        Override commit SHA (default: current HEAD)
-    --branch        Override branch name (default: current branch)
-    --message       Custom build message (default: auto-generated)
-    --help          Show this help message
+    --execute        Actually trigger the build (default: dry-run)
+    --pipeline       Buildkite pipeline slug (default: ${PIPELINE})
+    --commit         Override commit SHA (default: current HEAD)
+    --branch         Override branch name (default: current branch)
+    --message        Custom build message (default: auto-generated)
+    --torch-nightly  Also build and run the full suite against torch nightly
+                     (sets TORCH_NIGHTLY=1)
+    --help           Show this help message
 
 Prerequisites:
     - bk CLI installed: brew tap buildkite/buildkite && brew install buildkite/buildkite/bk
@@ -49,6 +52,7 @@ Examples:
     $(basename "$0")                        # Dry-run, show what would happen
     $(basename "$0") --execute              # Actually trigger the build
     $(basename "$0") --pipeline ci-shadow   # Dry-run with different pipeline
+    $(basename "$0") --torch-nightly        # Dry-run a full torch-nightly run
 EOF
     exit 1
 }
@@ -96,6 +100,10 @@ while [[ $# -gt 0 ]]; do
             MESSAGE="$2"
             shift 2
             ;;
+        --torch-nightly)
+            TORCH_NIGHTLY=true
+            shift
+            ;;
         --help|-h)
             usage
             ;;
@@ -171,11 +179,17 @@ if [[ $(echo "$REMOTE_BRANCHES" | wc -l) -gt 5 ]]; then
 fi
 echo ""
 
+# Environment variables passed to the build.
+BUILD_ENV=("RUN_ALL=1" "NIGHTLY=1")
+if [[ "$TORCH_NIGHTLY" == true ]]; then
+    BUILD_ENV+=("TORCH_NIGHTLY=1")
+fi
+
 log_info "Pipeline: ${PIPELINE}"
 log_info "Branch: ${BRANCH}"
 log_info "Commit: ${COMMIT}"
 log_info "Message: ${MESSAGE}"
-log_info "Environment: RUN_ALL=1, NIGHTLY=1"
+log_info "Environment: ${BUILD_ENV[*]}"
 echo ""
 
 # Build the command
@@ -187,9 +201,10 @@ CMD=(bk build create
     --commit "${COMMIT}"
     --branch "${BRANCH}"
     --message "${MESSAGE}"
-    --env "RUN_ALL=1"
-    --env "NIGHTLY=1"
 )
+for env_var in "${BUILD_ENV[@]}"; do
+    CMD+=(--env "${env_var}")
+done
 
 if [[ "$DRY_RUN" == true ]]; then
     echo "=========================================="
@@ -210,8 +225,14 @@ if [[ "$DRY_RUN" == true ]]; then
     echo "    --commit '$(escape_for_shell "${COMMIT}")' \\"
     echo "    --branch '$(escape_for_shell "${BRANCH}")' \\"
     echo "    --message '$(escape_for_shell "${MESSAGE}")' \\"
-    echo "    --env 'RUN_ALL=1' \\"
-    echo "    --env 'NIGHTLY=1'"
+    last_idx=$(( ${#BUILD_ENV[@]} - 1 ))
+    for i in "${!BUILD_ENV[@]}"; do
+        if [[ $i -eq $last_idx ]]; then
+            echo "    --env '$(escape_for_shell "${BUILD_ENV[$i]}")'"
+        else
+            echo "    --env '$(escape_for_shell "${BUILD_ENV[$i]}")' \\"
+        fi
+    done
     echo ""
     echo "=========================================="
     echo -e "${YELLOW}To actually trigger this build, run:${NC}"

.buildkite/scripts/upload-nightly-wheels.sh

@@ -6,8 +6,14 @@ set -ex
 # manylinux platform tag with auditwheel.
 # Index generation is handled separately by generate-and-upload-nightly-index.sh.
 
-# shellcheck source=lib/manylinux.sh
-source .buildkite/scripts/lib/manylinux.sh
+# auditwheel is Linux-only; macOS wheels already carry a valid tag, so skip the
+# manylinux retag for them.
+WHEEL_PLATFORM="${VLLM_WHEEL_PLATFORM:-linux}"
+
+if [[ "$WHEEL_PLATFORM" == "linux" ]]; then
+  # shellcheck source=lib/manylinux.sh
+  source .buildkite/scripts/lib/manylinux.sh
+fi
 
 BUCKET="vllm-wheels"
 SUBPATH=$BUILDKITE_COMMIT
@@ -27,8 +33,10 @@ wheel="${wheel_files[0]}"
 
 # ========= detect manylinux tag and rename ==========
 
-wheel="$(apply_manylinux_tag "$wheel")"
-echo "Renamed wheel to: $wheel"
+if [[ "$WHEEL_PLATFORM" == "linux" ]]; then
+  wheel="$(apply_manylinux_tag "$wheel")"
+  echo "Renamed wheel to: $wheel"
+fi
 
 # Extract the version from the wheel
 version=$(unzip -p "$wheel" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)

.buildkite/scripts/upload-rocm-wheels.sh

@@ -113,8 +113,8 @@ $PYTHON .buildkite/scripts/generate-nightly-index.py \
 echo "Uploading indices to $S3_COMMIT_PREFIX"
 aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "$S3_COMMIT_PREFIX"
 
-# Update rocm/nightly/ if on main branch and not a PR
-if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]] || [[ "$NIGHTLY" == "1" ]]; then
+# Only scheduled nightly builds should update the moving nightly index.
+if [[ "${NIGHTLY:-0}" == "1" ]]; then
     echo "Updating rocm/nightly/ index..."
     aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/rocm/nightly/"
 fi
@@ -147,7 +147,7 @@ echo ""
 echo "Install command (by commit):"
 echo "  pip install vllm --extra-index-url https://${BUCKET}.s3.amazonaws.com/$ROCM_SUBPATH/"
 echo ""
-if [[ "$BUILDKITE_BRANCH" == "main" ]] || [[ "$NIGHTLY" == "1" ]]; then
+if [[ "${NIGHTLY:-0}" == "1" ]]; then
     echo "Install command (nightly):"
     echo "  pip install vllm --extra-index-url https://${BUCKET}.s3.amazonaws.com/rocm/nightly/"
 fi

.buildkite/scripts/xpu/push-nightly-builds-xpu.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+set -ex
+
+ORIG_TAG_NAME="$BUILDKITE_COMMIT"
+REPO="vllm/vllm-openai-xpu"
+
+echo "Pushing original XPU tag ${ORIG_TAG_NAME}-xpu to nightly tags in ${REPO}"
+
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-x86_64-xpu
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-x86_64-xpu ${REPO}:nightly-x86_64
+docker push ${REPO}:nightly-x86_64
+
+docker manifest rm ${REPO}:nightly || true
+docker manifest rm ${REPO}:nightly-"$BUILDKITE_COMMIT" || true
+docker manifest create ${REPO}:nightly ${REPO}:nightly-x86_64 --amend
+docker manifest create ${REPO}:nightly-"$BUILDKITE_COMMIT" ${REPO}:nightly-x86_64 --amend
+docker manifest push ${REPO}:nightly
+docker manifest push ${REPO}:nightly-"$BUILDKITE_COMMIT"

.buildkite/test_areas/attention.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: V1 attention (H100-MI300)
   key: v1-attention-h100-mi300
-  timeout_in_minutes: 30
+  timeout_in_minutes: 85
   device: h100
   source_file_dependencies:
     - vllm/config/attention.py
@@ -12,11 +12,13 @@ steps:
     - vllm/v1/attention
     - tests/v1/attention
   commands:
-    - pytest -v -s v1/attention
+    - pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 2
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 70
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 125
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -30,7 +32,7 @@ steps:
 
 - label: V1 attention (B200)
   key: v1-attention-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 80
   device: b200-k8s
   source_file_dependencies:
     - vllm/config/attention.py
@@ -38,4 +40,5 @@ steps:
     - vllm/v1/attention
     - tests/v1/attention
   commands:
-    - pytest -v -s v1/attention
+    - pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 2

.buildkite/test_areas/basic_correctness.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Basic Correctness
   key: basic-correctness
-  timeout_in_minutes: 30
+  timeout_in_minutes: 68
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -18,7 +18,8 @@ steps:
   - pytest -v -s basic_correctness/test_cpu_offload.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 50
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 60
       depends_on:
       - image-build-amd

.buildkite/test_areas/benchmarks.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Benchmarks CLI Test
   key: benchmarks-cli-test
-  timeout_in_minutes: 20
+  timeout_in_minutes: 45
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -13,7 +13,9 @@ steps:
   - pytest -v -s benchmarks/
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 40
       depends_on:
       - image-build-amd
 
@@ -23,7 +25,7 @@ steps:
   num_gpus: 2
   optional: true
   working_dir: "/vllm-workspace/"
-  timeout_in_minutes: 10
+  timeout_in_minutes: 20
   source_file_dependencies:
   - benchmarks/attention_benchmarks/
   - vllm/v1/attention/

.buildkite/test_areas/compile.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Sequence Parallel Correctness Tests (2 GPUs)
   key: sequence-parallel-correctness-tests-2-gpus
-  timeout_in_minutes: 50
+  timeout_in_minutes: 80
   working_dir: "/vllm-workspace/"
   num_devices: 2
   source_file_dependencies:
@@ -19,7 +19,7 @@ steps:
 
 - label: Sequence Parallel Correctness Tests (2xH100)
   key: sequence-parallel-correctness-tests-2xh100
-  timeout_in_minutes: 50
+  timeout_in_minutes: 75
   working_dir: "/vllm-workspace/"
   device: h100
   optional: true
@@ -30,7 +30,7 @@ steps:
 
 - label: AsyncTP Correctness Tests (2xH100)
   key: asynctp-correctness-tests-2xh100
-  timeout_in_minutes: 50
+  timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
   device: h100
   optional: true
@@ -41,7 +41,7 @@ steps:
 
 - label: AsyncTP Correctness Tests (B200)
   key: asynctp-correctness-tests-b200
-  timeout_in_minutes: 50
+  timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
   device: b200-k8s
   optional: true
@@ -52,7 +52,7 @@ steps:
 
 - label: Distributed Compile Unit Tests (2xH100)
   key: distributed-compile-unit-tests-2xh100
-  timeout_in_minutes: 20
+  timeout_in_minutes: 45
   working_dir: "/vllm-workspace/"
   device: h100
   num_devices: 2
@@ -66,7 +66,7 @@ steps:
 
 - label: Fusion and Compile Unit Tests (2xB200)
   key: fusion-and-compile-unit-tests-2xb200
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
   device: b200-k8s
   source_file_dependencies:
@@ -96,7 +96,7 @@ steps:
 
 - label: Fusion E2E Quick (H100)
   key: fusion-e2e-quick-h100
-  timeout_in_minutes: 15
+  timeout_in_minutes: 25
   working_dir: "/vllm-workspace/"
   device: h100
   num_devices: 1
@@ -115,7 +115,7 @@ steps:
 
 - label: Fusion E2E Config Sweep (H100)
   key: fusion-e2e-config-sweep-h100
-  timeout_in_minutes: 30
+  timeout_in_minutes: 25
   working_dir: "/vllm-workspace/"
   device: h100
   num_devices: 1
@@ -149,7 +149,7 @@ steps:
 
 - label: Fusion E2E TP2 Quick (H100)
   key: fusion-e2e-tp2-quick-h100
-  timeout_in_minutes: 20
+  timeout_in_minutes: 35
   working_dir: "/vllm-workspace/"
   device: h100
   num_devices: 2
@@ -167,7 +167,7 @@ steps:
 
 - label: Fusion E2E TP2 AR-RMS Config Sweep (H100)
   key: fusion-e2e-tp2-ar-rms-config-sweep-h100
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
   device: h100
   num_devices: 2
@@ -207,7 +207,7 @@ steps:
 
 - label: Fusion E2E TP2 (B200)
   key: fusion-e2e-tp2-b200
-  timeout_in_minutes: 20
+  timeout_in_minutes: 45
   working_dir: "/vllm-workspace/"
   device: b200-k8s
   num_devices: 2

.buildkite/test_areas/cuda.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Platform Tests
   key: platform-tests
-  timeout_in_minutes: 15
+  timeout_in_minutes: 20
   device: h200_18gb
   source_file_dependencies:
   - vllm/envs.py
@@ -18,14 +18,18 @@ steps:
     - pytest -v -s cuda/test_platform_no_cuda_init.py
 
 - label: Cudagraph
+  device: h200_35gb
   key: cudagraph
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
   source_file_dependencies:
   - tests/v1/cudagraph
   - vllm/v1/cudagraph_dispatcher.py
   - vllm/config/compilation.py
   - vllm/compilation
+  - vllm/v1/worker/encoder_cudagraph.py
+  - vllm/v1/worker/encoder_cudagraph_defs.py
   commands:
     - pytest -v -s v1/cudagraph/test_cudagraph_dispatch.py
     - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
-    - pytest -v -s v1/cudagraph/test_breakable_cudagraph.py
+    - pytest -v -s v1/cudagraph/test_breakable_cudagraph.py
+    - pytest -v -s v1/cudagraph/test_encoder_cudagraph.py

.buildkite/test_areas/disaggregated.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Distributed NixlConnector PD accuracy (4 GPUs)
   key: distributed-nixlconnector-pd-accuracy-4-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 55
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
@@ -15,8 +15,9 @@ steps:
     - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
   mirror:
     amd:
+      dind: false
       device: mi300_4
-      timeout_in_minutes: 110
+      timeout_in_minutes: 60
       depends_on:
         - image-build-amd
       source_file_dependencies:
@@ -29,7 +30,7 @@ steps:
 
 - label: Distributed FlashInfer NixlConnector PD accuracy (4 GPUs)
   key: distributed-flashinfer-nixlconnector-pd-accuracy-4-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 55
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
@@ -65,8 +66,9 @@ steps:
     - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
   mirror:
     amd:
+      dind: false
       device: mi300_4
-      timeout_in_minutes: 50
+      timeout_in_minutes: 40
       depends_on:
         - image-build-amd
       source_file_dependencies:
@@ -79,7 +81,7 @@ steps:
 
 - label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
   key: crosslayer-kv-layout-distributed-nixlconnector-pd-accuracy-tests-4-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 55
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
@@ -90,8 +92,9 @@ steps:
     - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
   mirror:
     amd:
+      dind: false
       device: mi300_4
-      timeout_in_minutes: 110
+      timeout_in_minutes: 60
       depends_on:
         - image-build-amd
       source_file_dependencies:
@@ -104,7 +107,7 @@ steps:
 
 - label: Hybrid SSM NixlConnector PD accuracy tests (4 GPUs)
   key: hybrid-ssm-nixlconnector-pd-accuracy-tests-4-gpus
-  timeout_in_minutes: 25
+  timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
@@ -115,8 +118,9 @@ steps:
     - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
   mirror:
     amd:
+      dind: false
       device: mi300_4
-      timeout_in_minutes: 60
+      timeout_in_minutes: 55
       depends_on:
         - image-build-amd
       source_file_dependencies:
@@ -143,7 +147,7 @@ steps:
 
 - label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
   key: multiconnector-nixl-offloading-pd-accuracy-2-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:
@@ -158,7 +162,7 @@ steps:
 
 - label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
   key: nixlconnector-pd-spec-decode-acceptance-2-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   device: a100
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
@@ -171,8 +175,9 @@ steps:
     - bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
   mirror:
     amd:
+      dind: false
       device: mi300_2
-      timeout_in_minutes: 60
+      timeout_in_minutes: 45
       depends_on:
         - image-build-amd
       source_file_dependencies:
@@ -182,11 +187,11 @@ steps:
         - vllm/platforms/rocm.py
       commands:
         - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-        - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
+        - KV_CACHE_MEMORY_BYTES=8G ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
 
 - label: MultiConnector (Nixl+Offloading) PD edge cases (2 GPUs)
   key: multiconnector-nixl-offloading-pd-edge-cases-2-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 25
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:
@@ -198,3 +203,25 @@ steps:
   commands:
     - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh
+
+# P TP 4 - D DPEP 4 test case for DSv4-Flash
+- label: DSv4-Flash Disaggregated DP EP
+  key: dsv4-flash-disaggregated
+  timeout_in_minutes: 60
+  device: h200
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 8
+  env:
+    ENABLE_HMA_FLAG: "1"
+    DP_EP: "1"
+    GPU_MEMORY_UTILIZATION: "0.85"
+    PREFILLER_TP_SIZE: "4"
+    DECODER_TP_SIZE: "4"
+    PREFILL_BLOCK_SIZE: "256"
+    DECODE_BLOCK_SIZE: "256"
+    MODEL_NAMES: "deepseek-ai/DeepSeek-V4-Flash"
+    VLLM_SERVE_EXTRA_ARGS: "--trust-remote-code,--kv-cache-dtype,fp8"
+  commands:
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+    - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh

.buildkite/test_areas/distributed.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Distributed Comm Ops
   key: distributed-comm-ops
-  timeout_in_minutes: 20
+  timeout_in_minutes: 25
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:
@@ -18,7 +18,7 @@ steps:
 
 - label: Distributed DP Tests (2 GPUs)
   key: distributed-dp-tests-2-gpus
-  timeout_in_minutes: 20
+  timeout_in_minutes: 35
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:
@@ -39,7 +39,9 @@ steps:
   - DP_SIZE=2 pytest -v -s entrypoints/openai/test_multi_api_servers.py
   mirror:
     amd:
+      dind: false
       device: mi300_2
+      timeout_in_minutes: 45
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -55,7 +57,7 @@ steps:
 
 - label: Distributed Compile + RPC Tests (2 GPUs)
   key: distributed-compile-rpc-tests-2-gpus
-  timeout_in_minutes: 20
+  timeout_in_minutes: 65
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:
@@ -78,7 +80,7 @@ steps:
 
 - label: Distributed Torchrun + Shutdown Tests (2 GPUs)
   key: distributed-torchrun-shutdown-tests-2-gpus
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:
@@ -133,7 +135,7 @@ steps:
 
 - label: Distributed DP Tests (4 GPUs)
   key: distributed-dp-tests-4-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
@@ -154,7 +156,7 @@ steps:
 
 - label: Distributed Compile + Comm (4 GPUs)
   key: distributed-compile-comm-4-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 70
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
@@ -176,7 +178,7 @@ steps:
 
 - label: Distributed Tests (8xH100)
   key: distributed-tests-8xh100
-  timeout_in_minutes: 10
+  timeout_in_minutes: 20
   device: h100
   num_devices: 8
   working_dir: "/vllm-workspace/tests"
@@ -212,7 +214,7 @@ steps:
 
 - label: Distributed Tests (2xH100-2xMI300)
   key: distributed-tests-2xh100-2xmi300
-  timeout_in_minutes: 15
+  timeout_in_minutes: 30
   device: h100
   optional: true
   working_dir: "/vllm-workspace/"
@@ -259,7 +261,7 @@ steps:
 
 - label: Pipeline + Context Parallelism (4 GPUs)
   key: pipeline-context-parallelism-4-gpus
-  timeout_in_minutes: 60
+  timeout_in_minutes: 55
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
@@ -274,7 +276,7 @@ steps:
 
 - label: RayExecutorV2 (4 GPUs)
   key: rayexecutorv2-4-gpus
-  timeout_in_minutes: 60
+  timeout_in_minutes: 45
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:

.buildkite/test_areas/docker.yaml

@@ -3,7 +3,7 @@ depends_on:
   - image-build-cpu
 steps:
 - label: Docker Build Metadata
-  timeout_in_minutes: 10
+  timeout_in_minutes: 20
   device: cpu-small
   source_file_dependencies:
     - .buildkite/release-pipeline.yaml

.buildkite/test_areas/e2e_integration.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: DeepSeek V2-Lite Sync EPLB Accuracy (4xH100)
   key: deepseek-v2-lite-sync-eplb-accuracy-4xh100
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
   device: h100
   optional: true
   num_devices: 4
@@ -14,7 +14,7 @@ steps:
 
 - label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (4xH100)
   key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy-4xh100
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
   device: h100
   optional: true
   num_devices: 4
@@ -24,7 +24,7 @@ steps:
 
 - label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (2xB200)
   key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy-2xb200
-  timeout_in_minutes: 60
+  timeout_in_minutes: 20
   device: b200-k8s
   optional: true
   num_devices: 2
@@ -34,7 +34,7 @@ steps:
 
 - label: Qwen3-30B-A3B-FP8 DP4 Async EPLB Accuracy
   key: qwen3-30b-a3b-fp8-dp4-async-eplb-accuracy
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
   device: h100
   optional: true
   num_devices: 4
@@ -44,7 +44,7 @@ steps:
 
 - label: DeepSeek V2-Lite Prefetch Offload Accuracy (H100)
   key: deepseek-v2-lite-prefetch-offload-accuracy-h100
-  timeout_in_minutes: 60
+  timeout_in_minutes: 20
   device: h100
   optional: true
   num_devices: 1

.buildkite/test_areas/engine.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Engine
   key: engine
-  timeout_in_minutes: 15
+  timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:
   - vllm/compilation/
@@ -28,14 +28,15 @@ steps:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py test_jit_monitor.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 60
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 40
       depends_on:
       - image-build-amd
 
 - label: Engine (1 GPU)
   key: engine-1-gpu
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   source_file_dependencies:
     - vllm/v1/engine/
     - tests/v1/engine/
@@ -44,14 +45,14 @@ steps:
     - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 40
+      device: mi250_1
+      timeout_in_minutes: 45
       depends_on:
       - image-build-amd
 
 - label: e2e Scheduling (1 GPU)
   key: e2e-scheduling-1-gpu
-  timeout_in_minutes: 30
+  timeout_in_minutes: 53
   device: h200_18gb
   source_file_dependencies:
     - vllm/v1/
@@ -60,15 +61,15 @@ steps:
     - pytest -v -s v1/e2e/general/test_async_scheduling.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 60
+      device: mi250_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
 
 - label: e2e Core (1 GPU)
   device: h200_35gb
   key: e2e-core-1-gpu
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
   source_file_dependencies:
     - vllm/v1/
     - tests/v1/e2e/general/
@@ -76,8 +77,8 @@ steps:
     - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 35
+      device: mi250_1
+      timeout_in_minutes: 50
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -87,7 +88,7 @@ steps:
 
 - label: V1 e2e (2 GPUs)
   key: v1-e2e-2-gpus
-  timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
+  timeout_in_minutes: 25 # TODO: Fix timeout after we have more confidence in the test stability
   optional: true
   num_devices: 2
   source_file_dependencies:
@@ -114,13 +115,15 @@ steps:
     - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
   mirror:
     amd:
+      dind: false
       device: mi300_2
+      timeout_in_minutes: 30
       depends_on:
       - image-build-amd
 
 - label: V1 e2e (4 GPUs)
   key: v1-e2e-4-gpus
-  timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
+  timeout_in_minutes: 20 # TODO: Fix timeout after we have more confidence in the test stability
   optional: true
   num_devices: 4
   source_file_dependencies:
@@ -148,7 +151,7 @@ steps:
 
 - label: V1 e2e (4xH100)
   key: v1-e2e-4xh100
-  timeout_in_minutes: 60
+  timeout_in_minutes: 35
   device: h100
   num_devices: 4
   optional: true

.buildkite/test_areas/entrypoints.yaml

@@ -3,8 +3,9 @@ depends_on:
   - image-build
 steps:
 - label: Entrypoints Unit Tests
+  device: h200_35gb
   key: entrypoints-unit-tests
-  timeout_in_minutes: 10
+  timeout_in_minutes: 25
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/entrypoints
@@ -15,8 +16,9 @@ steps:
   - pytest -v -s entrypoints/weight_transfer
 
 - label: Entrypoints Integration (LLM)
+  device: h200_35gb
   key: entrypoints-integration-llm
-  timeout_in_minutes: 40
+  timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -28,16 +30,16 @@ steps:
   - pytest -v -s entrypoints/llm/offline_mode # Needs to avoid interference with other tests
   mirror:
     amd:
-      device: mi325_1
-      # TODO(akaratza): Test after Torch >= 2.12 bump
-      soft_fail: true
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (API Server)
   key: entrypoints-integration-api-server
   device: h200_35gb
-  timeout_in_minutes: 130
+  timeout_in_minutes: 75
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -50,13 +52,16 @@ steps:
   - pytest -v -s entrypoints/scale_out
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (API Server OpenAI - Part 1)
+  device: h200_35gb
   key: entrypoints-integration-api-server-openai-part-1
-  timeout_in_minutes: 50
+  timeout_in_minutes: 68
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -67,14 +72,16 @@ steps:
   - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/correctness
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 80
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (API Server OpenAI - Part 2)
+  device: h200_35gb
   key: entrypoints-integration-api-server-openai-part-2
-  timeout_in_minutes: 50
+  timeout_in_minutes: 83
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -86,12 +93,14 @@ steps:
   - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 80
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 70
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (API Server Generate)
+  device: h200_35gb
   key: entrypoints-integration-api-server-generate
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
@@ -108,12 +117,14 @@ steps:
   - pytest -v -s entrypoints/anthropic
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 60
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (Responses API)
+  device: h200_35gb
   key: entrypoints-integration-responses-api
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
@@ -126,7 +137,7 @@ steps:
 - label: Entrypoints Integration (Speech to Text)
   device: h200_35gb
   key: entrypoints-integration-speech_to_text
-  timeout_in_minutes: 50
+  timeout_in_minutes: 45
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -138,7 +149,7 @@ steps:
 - label: Entrypoints Integration (Multimodal)
   device: h200_35gb
   key: entrypoints-integration-multimodal
-  timeout_in_minutes: 50
+  timeout_in_minutes: 45
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -148,8 +159,9 @@ steps:
   - pytest -v -s entrypoints/multimodal
 
 - label: Entrypoints Integration (Pooling)
+  device: h200_35gb
   key: entrypoints-integration-pooling
-  timeout_in_minutes: 50
+  timeout_in_minutes: 75
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -160,7 +172,7 @@ steps:
 
 - label: OpenAI API Correctness
   key: openai-api-correctness
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
   device: h200_18gb
   source_file_dependencies:
   - csrc/
@@ -169,7 +181,9 @@ steps:
   - pytest -s entrypoints/openai/correctness/
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 30
       depends_on:
       - image-build-amd
       source_file_dependencies:

.buildkite/test_areas/expert_parallelism.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: EPLB Algorithm
   key: eplb-algorithm
-  timeout_in_minutes: 15
+  timeout_in_minutes: 20
   device: h200_18gb
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -16,7 +16,9 @@ steps:
   - pytest -v -s distributed/test_eplb_utils.py
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 30
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -27,7 +29,7 @@ steps:
 
 - label: EPLB Execution # 17min
   key: eplb-execution
-  timeout_in_minutes: 27
+  timeout_in_minutes: 25
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:
@@ -39,7 +41,7 @@ steps:
 
 - label: Elastic EP Scaling Test
   key: elastic-ep-scaling-test
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
   device: h100
   working_dir: "/vllm-workspace/tests"
   num_devices: 4

.buildkite/test_areas/kernels.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: vLLM IR Tests
   key: vllm-ir-tests
-  timeout_in_minutes: 10
+  timeout_in_minutes: 35
   device: h200_18gb
   working_dir: "/vllm-workspace/"
   source_file_dependencies:
@@ -15,19 +15,21 @@ steps:
     - pytest -v -s tests/kernels/ir
 
 - label: Kernels Core Operation Test
+  device: h200_35gb
   key: kernels-core-operation-test
-  timeout_in_minutes: 75
+  timeout_in_minutes: 120
   source_file_dependencies:
   - csrc/
   - tests/kernels/core
   - tests/kernels/test_concat_mla_q.py
   - tests/kernels/test_fused_qk_norm_rope_gate.py
   commands:
-    - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py  kernels/test_concat_mla_q.py kernels/test_fused_qk_norm_rope_gate.py
+    - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py kernels/test_concat_mla_q.py kernels/test_fused_qk_norm_rope_gate.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 3
 
 - label: Kernels MiniMax Reduce RMS Test (2 GPUs)
   key: kernels-minimax-reduce-rms-test-2-gpus
-  timeout_in_minutes: 15
+  timeout_in_minutes: 20
   num_devices: 2
   device: h100
   source_file_dependencies:
@@ -41,7 +43,7 @@ steps:
 
 - label: Deepseek V4 Kernel Test (H100)
   key: deepseek-v4-kernel-test-h100
-  timeout_in_minutes: 15
+  timeout_in_minutes: 30
   device: h100
   source_file_dependencies:
   - csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
@@ -54,7 +56,7 @@ steps:
 
 - label: Deepseek V4 Kernel Test (B200)
   key: deepseek-v4-kernel-test-b200
-  timeout_in_minutes: 15
+  timeout_in_minutes: 20
   device: b200-k8s
   source_file_dependencies:
   - csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
@@ -65,7 +67,7 @@ steps:
 
 - label: Kernels Attention Test %N
   key: kernels-attention-test
-  timeout_in_minutes: 35
+  timeout_in_minutes: 65
   source_file_dependencies:
   - csrc/attention/
   - vllm/v1/attention
@@ -78,8 +80,9 @@ steps:
   parallelism: 2
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 55
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 90
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -106,7 +109,7 @@ steps:
 
 - label: Kernels Quantization Test %N
   key: kernels-quantization-test
-  timeout_in_minutes: 90
+  timeout_in_minutes: 60
   source_file_dependencies:
   - csrc/quantization/
   - vllm/model_executor/layers/quantization
@@ -116,7 +119,9 @@ steps:
   parallelism: 2
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 120
       source_file_dependencies:
       - csrc/quantization/
       - vllm/model_executor/layers/quantization
@@ -131,7 +136,7 @@ steps:
 
 - label: Kernels MoE Test %N
   key: kernels-moe-test
-  timeout_in_minutes: 25
+  timeout_in_minutes: 50
   source_file_dependencies:
   - csrc/quantization/cutlass_w8a8/moe/
   - csrc/moe/
@@ -146,8 +151,9 @@ steps:
   parallelism: 5
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 50
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
       source_file_dependencies:
       - csrc/quantization/cutlass_w8a8/moe/
       - csrc/moe/
@@ -162,8 +168,9 @@ steps:
       - image-build-amd
 
 - label: Kernels Mamba Test
+  device: h200_35gb
   key: kernels-mamba-test
-  timeout_in_minutes: 45
+  timeout_in_minutes: 40
   source_file_dependencies:
   - csrc/mamba/
   - tests/kernels/mamba
@@ -172,19 +179,19 @@ steps:
     - pytest -v -s kernels/mamba
 
 - label: Kernels KDA Test
-  timeout_in_minutes: 20
+  timeout_in_minutes: 25
   device: h200_18gb
   source_file_dependencies:
-  - vllm/model_executor/layers/fla/ops/kda.py
-  - vllm/model_executor/layers/fla/ops/chunk_delta_h.py
-  - vllm/model_executor/layers/fla/ops/l2norm.py
+  - vllm/third_party/flash_linear_attention/ops/kda.py
+  - vllm/third_party/flash_linear_attention/ops/chunk_delta_h.py
+  - vllm/third_party/flash_linear_attention/ops/l2norm.py
   - tests/kernels/test_kda.py
   commands:
     - pytest -v -s kernels/test_kda.py
 
 - label: Kernels DeepGEMM Test (H100)
   key: kernels-deepgemm-test-h100
-  timeout_in_minutes: 45
+  timeout_in_minutes: 35
   device: h100
   num_devices: 1
   source_file_dependencies:
@@ -211,7 +218,7 @@ steps:
 
 - label: Kernels (B200)
   key: kernels-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 80
   working_dir: "/vllm-workspace/"
   device: b200-k8s
   # optional: true
@@ -231,6 +238,15 @@ steps:
   - vllm/v1/attention/backends/mla/flashinfer_mla.py
   - vllm/v1/attention/selector.py
   - vllm/platforms/cuda.py
+  - vllm/model_executor/kernels/linear/cute_dsl/ll_bf16.py
+  - vllm/model_executor/kernels/linear/cute_dsl/_ll_bf16_dotprod.py
+  - vllm/model_executor/kernels/linear/cute_dsl/_ll_bf16_splitk.py
+  - vllm/cute_utils/
+  - vllm/model_executor/layers/mamba/ops/gdn_chunk_cutedsl/
+  - vllm/model_executor/layers/fused_moe/router/bf16x3_router_gemm_cutedsl.py
+  - tests/kernels/mamba/test_gdn_prefill_cutedsl.py
+  - tests/kernels/test_bf16x3_router_gemm_cutedsl.py
+  - tests/kernels/test_ll_bf16_gemm.py
   - tests/kernels/test_top_k_per_row.py
   commands:
     - nvidia-smi
@@ -259,24 +275,28 @@ steps:
     - pytest -v -s tests/kernels/moe/test_flashinfer_moe.py
     - pytest -v -s tests/kernels/moe/test_trtllm_nvfp4_moe.py
     - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
+    - pytest -v -s tests/kernels/mamba/test_gdn_prefill_cutedsl.py
+    - pytest -v -s tests/kernels/test_bf16x3_router_gemm_cutedsl.py
+    - pytest -v -s tests/kernels/test_ll_bf16_gemm.py
     # e2e
     - pytest -v -s tests/models/quantization/test_nvfp4.py
 
 - label: Kernels Helion Test
   key: kernels-helion-test
-  timeout_in_minutes: 30
+  timeout_in_minutes: 115
   device: h100
   source_file_dependencies:
   - vllm/utils/import_utils.py
   - tests/kernels/helion/
   commands:
     - pip install helion==1.1.0
-    - pytest -v -s kernels/helion/
+    - pytest -v -s kernels/helion/ --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 2
 
  
 - label: Kernels FP8 MoE Test (1xH100)
   key: kernels-fp8-moe-test-1xh100
-  timeout_in_minutes: 90
+  timeout_in_minutes: 40
   device: h100
   num_devices: 1
   optional: true
@@ -293,7 +313,7 @@ steps:
 
 - label: Kernels FP8 MoE Test (2xH100)
   key: kernels-fp8-moe-test-2xh100
-  timeout_in_minutes: 90
+  timeout_in_minutes: 45
   device: h100
   num_devices: 2
   optional: true
@@ -303,7 +323,7 @@ steps:
 
 - label: Kernels Fp4 MoE Test (B200)
   key: kernels-fp4-moe-test-b200
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
   device: b200-k8s
   num_devices: 1
   optional: true
@@ -316,7 +336,7 @@ steps:
 
 - label: Kernels FusedMoE Layer Test (2 H100s)
   key: kernels-fusedmoe-layer-test-2-h100s
-  timeout_in_minutes: 90
+  timeout_in_minutes: 30
   device: h100
   num_devices: 2
   source_file_dependencies:

.buildkite/test_areas/lm_eval.yaml

@@ -5,7 +5,7 @@ steps:
 - label: LM Eval Small Models
   device: h200_35gb
   key: lm-eval-small-models
-  timeout_in_minutes: 75
+  timeout_in_minutes: 45
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
@@ -14,8 +14,9 @@ steps:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 55
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 45
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -56,7 +57,7 @@ steps:
 
 - label: LM Eval Small Models (1xB200)
   key: lm-eval-small-models-1xb200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 50
   device: b200-k8s
   optional: true
   source_file_dependencies:
@@ -78,9 +79,31 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small-tp.txt
 
+- label: LM Eval PCP (4xB200)
+  key: lm-eval-pcp-4xb200
+  timeout_in_minutes: 360
+  device: b200-k8s
+  num_devices: 4
+  optional: true
+  source_file_dependencies:
+  - csrc/
+  - tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP2-PCP2-EP.yaml
+  - tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP1-PCP4-EP.yaml
+  - tests/evals/gsm8k/configs/models-pcp.txt
+  - vllm/model_executor/layers/quantization
+  - vllm/config/parallel.py
+  - vllm/distributed/parallel_state.py
+  - vllm/model_executor/layers/attention/mla_attention.py
+  - vllm/model_executor/layers/attention/pcp.py
+  - vllm/v1/worker/gpu/model_runner.py
+  - vllm/v1/worker/gpu/pcp_manager.py
+  autorun_on_main: true
+  commands:
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-pcp.txt
+
 - label: LM Eval Large Models EP (2xB200)
   key: lm-eval-large-models-ep-2xb200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 60
   device: b200-k8s
   optional: true
   num_devices: 2
@@ -92,7 +115,7 @@ steps:
 
 - label: LM Eval Qwen3.5 Models (2xB200)
   key: lm-eval-qwen3-5-models-2xb200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 45
   device: b200-k8s
   optional: true
   num_devices: 2
@@ -103,13 +126,13 @@ steps:
   - vllm/transformers_utils/configs/qwen3_5_moe.py
   - vllm/model_executor/models/qwen3_next.py
   - vllm/model_executor/models/qwen3_next_mtp.py
-  - vllm/model_executor/layers/fla/ops/
+  - vllm/third_party/flash_linear_attention/ops/
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-blackwell.txt
 
 - label: LM Eval Large Models (8xH200)
   key: lm-eval-large-models-8xh200
-  timeout_in_minutes: 60
+  timeout_in_minutes: 50
   device: h200
   optional: true
   num_devices: 8
@@ -117,8 +140,9 @@ steps:
     - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200.txt
   mirror:
     amd:
+      dind: false
       device: mi300_8
-      timeout_in_minutes: 180
+      timeout_in_minutes: 40
       depends_on:
       - image-build-amd
       commands:
@@ -152,7 +176,7 @@ steps:
 
 - label: LM Eval Humming f16 (A100 - TEMPORARY)
   key: lm-eval-humming-f16-a100
-  timeout_in_minutes: 120
+  timeout_in_minutes: 75
   device: a100
   optional: true
   num_devices: 1
@@ -167,7 +191,7 @@ steps:
 
 - label: LM Eval Humming Act int8 (A100 - TEMPORARY)
   key: lm-eval-humming-act-a100
-  timeout_in_minutes: 120
+  timeout_in_minutes: 45
   device: a100
   optional: true
   num_devices: 1
@@ -182,7 +206,7 @@ steps:
 
 - label: LM Eval Humming f16 (H100 - TEMPORARY)
   key: lm-eval-humming-f16-h100
-  timeout_in_minutes: 120
+  timeout_in_minutes: 70
   device: h100
   optional: true
   num_devices: 1
@@ -197,7 +221,7 @@ steps:
 
 - label: LM Eval Humming Act fp8/int8 (H100 - TEMPORARY)
   key: lm-eval-humming-act-h100
-  timeout_in_minutes: 120
+  timeout_in_minutes: 70
   device: h100
   optional: true
   num_devices: 1
@@ -213,7 +237,7 @@ steps:
 
 - label: LM Eval Humming f16 (B200 - TEMPORARY)
   key: lm-eval-humming-f16-b200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 50
   device: b200-k8s
   optional: true
   num_devices: 1
@@ -228,7 +252,7 @@ steps:
 
 - label: LM Eval Humming Act fp8/int8 (B200 - TEMPORARY)
   key: lm-eval-humming-act-b200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 50
   device: b200-k8s
   optional: true
   num_devices: 1
@@ -244,7 +268,7 @@ steps:
 
 - label: LM Eval TurboQuant KV Cache
   key: lm-eval-turboquant-kv-cache
-  timeout_in_minutes: 75
+  timeout_in_minutes: 55
   device: h200_18gb
   source_file_dependencies:
   - vllm/model_executor/layers/quantization/turboquant/
@@ -256,7 +280,7 @@ steps:
 
 - label: GPQA Eval (GPT-OSS) (2xH100)
   key: gpqa-eval-gpt-oss-2xh100
-  timeout_in_minutes: 120
+  timeout_in_minutes: 35
   device: h100
   optional: true
   num_devices: 2
@@ -270,7 +294,7 @@ steps:
 
 - label: GPQA Eval (GPT-OSS) (2xB200)
   key: gpqa-eval-gpt-oss-2xb200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 30
   device: b200-k8s
   optional: true
   num_devices: 2
@@ -284,7 +308,7 @@ steps:
 
 - label: GPQA Eval (GPT-OSS) (DGX Spark)
   key: gpqa-eval-gpt-oss-spark
-  timeout_in_minutes: 120
+  timeout_in_minutes: 35
   device: dgx-spark
   optional: true
   num_devices: 1
@@ -313,7 +337,7 @@ steps:
 
 - label: LM Eval KV-Offload (2xH100)
   key: kv-offload-medium
-  timeout_in_minutes: 60
+  timeout_in_minutes: 30
   device: h100
   num_devices: 2
   source_file_dependencies:
@@ -327,7 +351,7 @@ steps:
 
 - label: LM Eval KV-Offload (4xH100)
   key: kv-offload-large
-  timeout_in_minutes: 60
+  timeout_in_minutes: 40
   device: h100
   num_devices: 4
   source_file_dependencies:
@@ -341,7 +365,7 @@ steps:
 
 - label: MRCR Eval Small Models
   device: h200_35gb
-  timeout_in_minutes: 30
+  timeout_in_minutes: 25
   source_file_dependencies:
   - tests/evals/mrcr/
   commands:

.buildkite/test_areas/lora.yaml

@@ -5,7 +5,7 @@ steps:
 - label: LoRA %N
   device: h200_35gb
   key: lora
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
   source_file_dependencies:
   - vllm/lora
   - tests/lora
@@ -14,9 +14,10 @@ steps:
   parallelism: 4
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
       working_dir: "/vllm-workspace/tests"
-      timeout_in_minutes: 60
+      timeout_in_minutes: 85
       source_file_dependencies:
       - vllm/lora
       - tests/lora
@@ -27,7 +28,7 @@ steps:
 
 - label: LoRA TP (Distributed)
   key: lora-tp-distributed
-  timeout_in_minutes: 30
+  timeout_in_minutes: 60
   num_devices: 4
   source_file_dependencies:
   - vllm/lora
@@ -46,4 +47,4 @@ steps:
     - pytest -v -s -x lora/test_qwen3_with_multi_loras.py
     - pytest -v -s -x lora/test_olmoe_tp.py
     - pytest -v -s -x lora/test_gptoss_tp.py
-    - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
+    - pytest -v -s -x lora/test_qwen35_densemodel_lora.py

.buildkite/test_areas/misc.yaml

@@ -5,7 +5,7 @@ steps:
 - label: V1 Spec Decode
   device: h200_35gb
   key: v1-spec-decode
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
   source_file_dependencies:
     - vllm/config/
     - vllm/distributed/
@@ -23,14 +23,15 @@ steps:
     - pytest -v -s -m 'not slow_test' v1/spec_decode
   mirror:
     amd:
+      dind: false
       device: mi300_1
-      timeout_in_minutes: 65
+      timeout_in_minutes: 50
       depends_on:
       - image-build-amd
 
 - label: V1 Sample + Logits
   key: v1-sample-logits
-  timeout_in_minutes: 30
+  timeout_in_minutes: 83
   device: h200_18gb
   source_file_dependencies:
     - vllm/config/
@@ -58,13 +59,16 @@ steps:
     - pytest -v -s v1/test_outputs.py
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 70
       depends_on:
       - image-build-amd
 
 - label: V1 Core + KV + Metrics
+  device: h200_35gb
   key: v1-core-kv-metrics
-  timeout_in_minutes: 30
+  timeout_in_minutes: 80
   source_file_dependencies:
     - vllm/config/
     - vllm/distributed/
@@ -88,7 +92,9 @@ steps:
     - tests/v1/kv_offload
     - tests/v1/simple_kv_offload
     - tests/v1/worker
+    - tests/v1/streaming_input
     - tests/v1/kv_connector/unit
+    - tests/v1/ec_connector/unit
     - tests/v1/metrics
     - tests/entrypoints/openai/correctness/test_lmeval.py
   commands:
@@ -100,15 +106,18 @@ steps:
     - pytest -v -s v1/kv_offload
     - pytest -v -s v1/simple_kv_offload
     - pytest -v -s v1/worker
+    - pytest -v -s v1/streaming_input
     - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
+    - pytest -v -s -m 'not cpu_test' v1/ec_connector/unit
     - pytest -v -s -m 'not cpu_test' v1/metrics
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/vllm-project/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 60
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
       depends_on:
       - image-build-amd
 
@@ -139,6 +148,7 @@ steps:
     - pytest -v -s -m 'cpu_test' v1/core
     - pytest -v -s v1/structured_output
     - pytest -v -s v1/test_serial_utils.py
+    - pytest -v -s v1/cudagraph/test_cudagraph_manager.py
     - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
     - pytest -v -s -m 'cpu_test' v1/metrics
 
@@ -172,7 +182,7 @@ steps:
 
 - label: Regression
   key: regression
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:
   - vllm/config/
@@ -195,14 +205,14 @@ steps:
 - label: Examples
   device: h200_35gb
   key: examples
-  timeout_in_minutes: 45
+  timeout_in_minutes: 40
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
   - vllm/entrypoints
   - vllm/multimodal
   - examples/
   commands:
-    - pip install tensorizer # for tensorizer test
+    - pip install --no-deps tensorizer # for tensorizer test
      # for basic
     - python3 basic/offline_inference/chat.py
     - python3 basic/offline_inference/generate.py --model facebook/opt-125m
@@ -226,7 +236,9 @@ steps:
     - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 75
       source_file_dependencies:
       - vllm/entrypoints
       - vllm/multimodal
@@ -237,7 +249,7 @@ steps:
 
 - label: Metrics, Tracing (2 GPUs)
   key: metrics-tracing-2-gpus
-  timeout_in_minutes: 20
+  timeout_in_minutes: 25
   num_devices: 2
   source_file_dependencies:
   - vllm/config/
@@ -262,10 +274,11 @@ steps:
   - pytest -v -s v1/tracing
   mirror:
     amd:
-      device: mi325_2
+      dind: false
+      device: mi300_2
+      timeout_in_minutes: 30
       depends_on:
       - image-build-amd
-      optional: true
 
 - label: Python-only Installation
   key: python-only-installation
@@ -280,8 +293,8 @@ steps:
   - bash standalone_tests/python_only_compile.sh
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 20
+      device: mi250_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -292,7 +305,7 @@ steps:
 - label: Async Engine, Inputs, Utils, Worker
   device: h200_35gb
   key: async-engine-inputs-utils-worker
-  timeout_in_minutes: 50
+  timeout_in_minutes: 25
   source_file_dependencies:
   - vllm/assets/
   - vllm/config/
@@ -319,7 +332,7 @@ steps:
   key: async-engine-inputs-utils-worker-config-cpu
   depends_on:
   - image-build-cpu
-  timeout_in_minutes: 30
+  timeout_in_minutes: 65
   source_file_dependencies:
   - vllm/assets/
   - vllm/config/
@@ -381,7 +394,7 @@ steps:
 
 - label: Batch Invariance (A100)
   key: batch-invariance-a100
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
   device: a100
   source_file_dependencies:
     - vllm/v1/attention
@@ -395,7 +408,7 @@ steps:
 
 - label: Batch Invariance (H100)
   key: batch-invariance-h100
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
   device: h100
   source_file_dependencies:
     - vllm/v1/attention
@@ -411,7 +424,7 @@ steps:
 
 - label: Batch Invariance (B200)
   key: batch-invariance-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 35
   device: b200-k8s
   source_file_dependencies:
     - vllm/v1/attention
@@ -430,7 +443,7 @@ steps:
 - label: Acceptance Length Test (Large Models) # optional
   device: h200_35gb
   key: acceptance-length-test-large-models
-  timeout_in_minutes: 25
+  timeout_in_minutes: 20
   gpu: h100
   optional: true
   num_gpus: 1

.buildkite/test_areas/model_executor.yaml

@@ -3,13 +3,16 @@ depends_on:
   - image-build
 steps:
 - label: Model Executor
+  device: h200_35gb
   key: model-executor
-  timeout_in_minutes: 35
+  timeout_in_minutes: 60
   source_file_dependencies:
   - 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
@@ -25,14 +28,18 @@ steps:
     - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py --timeout=900 --timeout-method=thread
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 60
       depends_on:
       - image-build-amd
       source_file_dependencies:
       - 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

.buildkite/test_areas/model_runner_v2.yaml

@@ -5,7 +5,7 @@ steps:
 - label: Model Runner V2 Core Tests
   device: h200_35gb
   key: model-runner-v2-core-tests
-  timeout_in_minutes: 45
+  timeout_in_minutes: 35
   source_file_dependencies:
   - vllm/v1/worker/gpu/
   - vllm/v1/worker/gpu_worker.py
@@ -27,7 +27,7 @@ steps:
 - label: Model Runner V2 Examples
   device: h200_35gb
   key: model-runner-v2-examples
-  timeout_in_minutes: 45
+  timeout_in_minutes: 35
   working_dir: "/vllm-workspace/examples"
   source_file_dependencies:
     - vllm/v1/worker/gpu/
@@ -41,7 +41,7 @@ steps:
   commands:
     - set -x
     - export VLLM_USE_V2_MODEL_RUNNER=1
-    - pip install tensorizer # for tensorizer test
+    - pip install --no-deps tensorizer # for tensorizer test
     - python3 basic/offline_inference/chat.py # for basic
     - python3 basic/offline_inference/generate.py --model facebook/opt-125m
     #- python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10  # TODO
@@ -63,7 +63,7 @@ steps:
 
 - label: Model Runner V2 Distributed (2 GPUs)
   key: model-runner-v2-distributed-2-gpus
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:
@@ -84,7 +84,7 @@ steps:
 
 - label: Model Runner V2 Pipeline Parallelism (4 GPUs)
   key: model-runner-v2-pipeline-parallelism-4-gpus
-  timeout_in_minutes: 60
+  timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:

.buildkite/test_areas/models_basic.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Basic Models Tests (Initialization)
   key: basic-models-tests-initialization
-  timeout_in_minutes: 45
+  timeout_in_minutes: 25
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -17,7 +17,7 @@ steps:
 - label: Basic Models Tests (Extra Initialization) %N
   device: h200_35gb
   key: basic-models-tests-extra-initialization
-  timeout_in_minutes: 45
+  timeout_in_minutes: 100
   source_file_dependencies:
   - vllm/model_executor/models/
   - tests/models/test_initialization.py
@@ -27,12 +27,12 @@ steps:
     # subset of supported models (the complement of the small subset in the above
     # test.) Also run if model initialization test file is modified
     - pytest -v -s models/test_initialization.py -k 'not test_can_initialize_small_subset' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
-  parallelism: 2
+  parallelism: 4
 
 - label: Basic Models Tests (Other)
   device: h200_35gb
   key: basic-models-tests-other
-  timeout_in_minutes: 45
+  timeout_in_minutes: 35
   source_file_dependencies:
   - vllm/
   - tests/models/test_terratorch.py
@@ -42,15 +42,30 @@ steps:
     - pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 50
       depends_on:
       - image-build-amd
 
+- label: Inkling Unit Tests (B200)
+  key: inkling-unit-tests-b200
+  timeout_in_minutes: 40
+  device: b200-k8s
+  source_file_dependencies:
+  - vllm/models/inkling/
+  - vllm/cute_utils/
+  - cmake/external_projects/tml_fa4.cmake
+  - tests/models/inkling/
+  commands:
+    # FA4 kernel tests require SM100; the suite skips them elsewhere.
+    - pytest -v -s models/inkling
+
 - label: Basic Models Test (Other CPU) # 5min
   key: basic-models-test-other-cpu
   depends_on:
   - image-build-cpu
-  timeout_in_minutes: 10
+  timeout_in_minutes: 20
   source_file_dependencies:
   - vllm/
   - tests/models/test_utils.py

.buildkite/test_areas/models_distributed.yaml

@@ -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: 60
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:

.buildkite/test_areas/models_language.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Language Models Tests (Standard)
   key: language-models-tests-standard
-  timeout_in_minutes: 25
+  timeout_in_minutes: 30
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -15,13 +15,16 @@ steps:
     - pytest -v -s models/language -m 'core_model and (not slow_test)'
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 45
       depends_on:
       - image-build-amd
 
 - label: Language Models Tests (Extra Standard) %N
+  device: h200_35gb
   key: language-models-tests-extra-standard
-  timeout_in_minutes: 45
+  timeout_in_minutes: 40
   source_file_dependencies:
   - vllm/model_executor/models/
   - tests/models/language/pooling/test_embedding.py
@@ -35,7 +38,9 @@ steps:
   parallelism: 2
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 40
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -49,25 +54,25 @@ steps:
       - tests/models/language/pooling/test_classification.py
       - vllm/_aiter_ops.py
       - vllm/platforms/rocm.py
-
 - label: Language Models Tests (Hybrid) %N
+  device: h200_35gb
   key: language-models-tests-hybrid
-  timeout_in_minutes: 75
+  timeout_in_minutes: 65
   source_file_dependencies:
   - vllm/
   - tests/models/language/generation
   commands:
     # Install fast path packages for testing against transformers
-    # Note: also needed to run plamo2 model in vLLM
     - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
-    # Shard hybrid language model tests
-    - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+    # Shard the hybrid language model tests that are numerically stable on Hopper.
+    - pytest -v -s models/language/generation -m hybrid_model -k 'not granite-4.0-tiny-preview' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
   parallelism: 2
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 90
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 60
       depends_on:
       - image-build-amd
       commands:
@@ -75,24 +80,37 @@ steps:
       - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
       - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
 
+# Granite 4 hybrid generation is sensitive to hardware-specific Triton SSD
+# autotuning (https://github.com/vllm-project/vllm/issues/25194). Keep this one
+# correctness test on L4 until its H200 output matches the Transformers reference.
+- label: Language Models Tests (Granite L4 Compatibility)
+  key: language-models-tests-granite-l4-compatibility
+  timeout_in_minutes: 65
+  source_file_dependencies:
+  - vllm/
+  - tests/models/language/generation
+  commands:
+    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
+    - pytest -v -s models/language/generation -m hybrid_model -k 'granite-4.0-tiny-preview'
+
 - label: Language Models Test (Extended Generation) # 80min
   device: h200_35gb
   key: language-models-test-extended-generation
-  timeout_in_minutes: 110
+  timeout_in_minutes: 65
   optional: true
   source_file_dependencies:
   - vllm/
   - tests/models/language/generation
   commands:
     # Install fast path packages for testing against transformers
-    # Note: also needed to run plamo2 model in vLLM
     - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
     - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 - label: Language Models Test (PPL)
   key: language-models-test-ppl
-  timeout_in_minutes: 110
+  timeout_in_minutes: 30
   device: h200_18gb
   optional: true
   source_file_dependencies:
@@ -101,10 +119,10 @@ steps:
   commands:
     - pytest -v -s models/language/generation_ppl_test
 
-- label: Language Models Test (Extended Pooling)  # 36min
+- label: Language Models Test (Extended Pooling)
   device: h200_35gb
   key: language-models-test-extended-pooling
-  timeout_in_minutes: 50
+  timeout_in_minutes: 120
   optional: true
   source_file_dependencies:
   - vllm/
@@ -113,14 +131,15 @@ steps:
     - pytest -v -s models/language/pooling -m 'not core_model'
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 100
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 95
       depends_on:
       - image-build-amd
 
 - label: Language Models Test (MTEB)
   key: language-models-test-mteb
-  timeout_in_minutes: 110
+  timeout_in_minutes: 68
   device: h200_18gb
   optional: true
   source_file_dependencies:

.buildkite/test_areas/models_multimodal.yaml

@@ -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: 68
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -14,13 +14,15 @@ steps:
     - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
       depends_on:
       - image-build-amd
 
 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
   key: multi-modal-models-standard-2-qwen3-gemma
-  timeout_in_minutes: 45
+  timeout_in_minutes: 75
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -31,14 +33,16 @@ steps:
     - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
 
 - label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
   device: h200_35gb
   key: multi-modal-models-standard-3-llava-qwen2-vl
-  timeout_in_minutes: 45
+  timeout_in_minutes: 40
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
@@ -47,14 +51,15 @@ steps:
     - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
   mirror:
     amd:
-      device: mi325_1
+      device: mi250_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
 
 - label: "Multi-Modal Models (Standard) 4: other + whisper"
   device: h200_35gb
   key: multi-modal-models-standard-4-other-whisper
-  timeout_in_minutes: 45
+  timeout_in_minutes: 75
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
@@ -65,26 +70,29 @@ steps:
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 50
       depends_on:
       - image-build-amd
 
-- label: Multi-Modal Processor (CPU)
+- label: Multi-Modal Processor (CPU) %N
   key: multi-modal-processor-cpu
   depends_on:
   - image-build-cpu
-  timeout_in_minutes: 60
+  timeout_in_minutes: 125
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
   - tests/models/registry.py
   device: cpu-medium
   commands:
-    - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py
+    - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+  parallelism: 4
 
 - label: Multi-Modal Processor # 44min
   key: multi-modal-processor
-  timeout_in_minutes: 60
+  timeout_in_minutes: 98
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -96,7 +104,7 @@ steps:
 - label: Multi-Modal Accuracy Eval (Small Models) # 50min
   device: h200_35gb
   key: multi-modal-accuracy-eval-small-models
-  timeout_in_minutes: 70
+  timeout_in_minutes: 30
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
   - vllm/multimodal/
@@ -106,7 +114,9 @@ steps:
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 35
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -117,6 +127,7 @@ steps:
       - vllm/model_executor/model_loader/
 
 - label: Multi-Modal Models (Extended Generation 1)
+  device: h200_35gb
   key: multi-modal-models-extended-generation-1
   optional: true
   source_file_dependencies:
@@ -128,7 +139,9 @@ steps:
     - pytest -v -s models/multimodal/test_mapping.py
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 90
       depends_on:
       - image-build-amd
 
@@ -163,7 +176,8 @@ steps:
     - pytest -v -s models/multimodal/pooling -m 'not core_model'
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
       timeout_in_minutes: 60
       depends_on:
       - image-build-amd

.buildkite/test_areas/plugins.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Plugin Tests (2 GPUs)
   key: plugin-tests-2-gpus
-  timeout_in_minutes: 60
+  timeout_in_minutes: 35
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:

.buildkite/test_areas/pytorch.yaml

@@ -5,7 +5,7 @@ steps:
 - label: PyTorch Compilation Unit Tests
   device: h200_35gb
   key: pytorch-compilation-unit-tests
-  timeout_in_minutes: 10
+  timeout_in_minutes: 150
   source_file_dependencies:
     - vllm/__init__.py
     - vllm/_aiter_ops.py
@@ -78,7 +78,7 @@ steps:
 
 - label: PyTorch Compilation Passes Unit Tests
   key: pytorch-compilation-passes-unit-tests
-  timeout_in_minutes: 20
+  timeout_in_minutes: 45
   source_file_dependencies:
     - vllm/__init__.py
     - vllm/_aiter_ops.py
@@ -107,16 +107,11 @@ steps:
     - tests/compile/passes
   commands:
   - pytest -s -v compile/passes --ignore compile/passes/distributed
-  mirror:
-    amd:
-      device: mi300_1
-      timeout_in_minutes: 180
-      depends_on:
-      - image-build-amd
 
 - label: PyTorch Fullgraph Smoke Test
+  device: h200_35gb
   key: pytorch-fullgraph-smoke-test
-  timeout_in_minutes: 35
+  timeout_in_minutes: 90
   source_file_dependencies:
   - vllm/__init__.py
   - vllm/_aiter_ops.py
@@ -148,11 +143,46 @@ steps:
   # as it is a heavy test that is covered in other steps.
   # Use `find` to launch multiple instances of pytest so that
   # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
-  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
+  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_cudagraph.py' -not -name 'test_full_graph.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
+
+# Hopper-only DeepSeek-V2-Lite cases in this file require two 29.3-GiB model
+# instances and cannot fit a 35GB MIG slice. L4 retains the original coverage:
+# those SM90 cases skip while the architecture-compatible cases still run.
+- label: PyTorch Fullgraph CUDAGraph (L4 Compatibility)
+  key: pytorch-fullgraph-cudagraph-l4-compatibility
+  timeout_in_minutes: 60
+  source_file_dependencies:
+  - vllm/__init__.py
+  - vllm/_aiter_ops.py
+  - vllm/_custom_ops.py
+  - vllm/compilation/
+  - vllm/config/
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/env_override.py
+  - vllm/envs.py
+  - vllm/forward_context.py
+  - vllm/inputs/
+  - vllm/ir/
+  - vllm/kernels/
+  - vllm/logger.py
+  - vllm/model_executor/
+  - vllm/multimodal/
+  - vllm/platforms/
+  - vllm/plugins/
+  - vllm/sampling_params.py
+  - vllm/sequence.py
+  - vllm/transformers_utils/
+  - vllm/triton_utils/
+  - vllm/utils/
+  - vllm/v1/
+  - tests/compile
+  commands:
+  - pytest -s -v compile/fullgraph/test_full_cudagraph.py
 
 - label: PyTorch Fullgraph
   key: pytorch-fullgraph
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
   device: h200_18gb
   source_file_dependencies:
   - vllm/__init__.py
@@ -197,7 +227,9 @@ steps:
   - bash standalone_tests/pytorch_nightly_dependency.sh
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 30
       depends_on:
       - image-build-amd
       source_file_dependencies:

.buildkite/test_areas/quantization.yaml

@@ -3,8 +3,11 @@ depends_on:
   - image-build
 steps:
 - label: Quantization
+  device: h200_35gb
   key: quantization
-  timeout_in_minutes: 90
+  timeout_in_minutes: 75
+  env:
+    VLLM_USE_V2_MODEL_RUNNER: "0"
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
@@ -19,11 +22,14 @@ steps:
   # TODO(jerryzh168): resolve the above comment
   - uv pip install --system torchao==0.17.0 --index-url https://download.pytorch.org/whl/cu130
   - uv pip install --system conch-triton-kernels
-  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
+  # The SM90-only checkpoint currently contains a removed weight_chan_scale
+  # parameter. It was not exercised by the previous L4 job.
+  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py -k 'not test_compressed_tensors_w4a8_fp8'
 
 - label: Quantized Fusions
+  device: h200_35gb
   key: quantized-fusions
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
   source_file_dependencies:
   - tests/fusion
   - vllm/model_executor/layers/fusion
@@ -35,7 +41,7 @@ steps:
 
 - label: Quantized MoE Test (B200)
   key: quantized-moe-test-b200
-  timeout_in_minutes: 60
+  timeout_in_minutes: 120
   working_dir: "/vllm-workspace/"
   device: b200-k8s
   source_file_dependencies:
@@ -52,8 +58,11 @@ steps:
     - pytest -s -v tests/quantization/test_blackwell_moe.py
 
 - label: Quantized Models Test
+  device: h200_35gb
   key: quantized-models-test
-  timeout_in_minutes: 60
+  timeout_in_minutes: 65
+  env:
+    VLLM_USE_V2_MODEL_RUNNER: "0"
   source_file_dependencies:
   - vllm/model_executor/layers/quantization
   - tests/models/quantization

.buildkite/test_areas/rust_frontend.yaml

@@ -3,7 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Rust Frontend OpenAI Coverage
-  timeout_in_minutes: 90
+  timeout_in_minutes: 30
   device: h200_18gb
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -38,7 +38,7 @@ steps:
   - pytest -v -s v1/sample/test_logprobs_e2e.py -k "test_prompt_logprobs_e2e_server"
 
 - label: Rust Frontend Serve/Admin Coverage
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
   device: h200_18gb
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -67,7 +67,7 @@ steps:
   - pytest -v -s entrypoints/serve/tokenize/test_tokenization.py -k "not tokenizer_info"
 
 - label: Rust Frontend Core Correctness
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
   device: h200_18gb
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -81,7 +81,8 @@ steps:
   - pytest -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
 - label: Rust Frontend Tool Use
-  timeout_in_minutes: 60
+  device: h200_35gb
+  timeout_in_minutes: 25
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - rust/
@@ -95,7 +96,7 @@ steps:
   - pytest -v -s tool_use --ignore=tool_use/mistral --models llama3.2 -k "not test_response_format_with_tool_choice_required and not test_parallel_tool_calls_false and not test_tool_call_and_choice"
 
 - label: Rust Frontend Distributed
-  timeout_in_minutes: 30
+  timeout_in_minutes: 25
   num_devices: 4
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:

.buildkite/test_areas/rust_frontend_cargo.yaml

@@ -4,7 +4,7 @@ steps:
 - label: Rust Frontend Cargo Style + Clippy
   key: rust-frontend-cargo-style-clippy
   depends_on: []
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
   device: cpu-medium
   no_plugin: true
   source_file_dependencies:
@@ -18,7 +18,7 @@ steps:
 - label: Rust Frontend Cargo Tests
   key: rust-frontend-cargo-tests
   depends_on: []
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
   device: cpu-medium
   no_plugin: true
   source_file_dependencies:

.buildkite/test_areas/samplers.yaml

@@ -5,7 +5,7 @@ steps:
 - label: Samplers Test
   device: h200_35gb
   key: samplers-test
-  timeout_in_minutes: 75
+  timeout_in_minutes: 40
   source_file_dependencies:
   - vllm/model_executor/layers
   - vllm/sampling_metadata.py
@@ -19,8 +19,18 @@ steps:
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
   mirror:
     amd:
-      device: mi325_1
+      device: mi250_1
+      timeout_in_minutes: 40
       depends_on:
       - image-build-amd
+      source_file_dependencies:
+      - vllm/model_executor/layers
+      - vllm/sampling_metadata.py
+      - vllm/v1/sample/
+      - vllm/entrypoints/generate/beam_search/
+      - tests/samplers
+      - tests/conftest.py
+      - vllm/_aiter_ops.py
+      - vllm/platforms/rocm.py
       commands:
       - pytest -v -s samplers

.buildkite/test_areas/spec_decode.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Spec Decode Eagle
   key: spec-decode-eagle
-  timeout_in_minutes: 30
+  timeout_in_minutes: 25
   device: h200_18gb
   source_file_dependencies:
     - vllm/v1/spec_decode/
@@ -14,8 +14,9 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 45
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -29,7 +30,7 @@ steps:
 
 - label: Spec Decode Eagle Nightly B200
   key: spec-decode-eagle-nightly-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 25
   device: b200-k8s
   optional: true
   source_file_dependencies:
@@ -41,7 +42,7 @@ steps:
 
 - label: Spec Decode Speculators + MTP
   key: spec-decode-speculators-mtp
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
   device: h200_18gb
   source_file_dependencies:
     - vllm/v1/spec_decode/
@@ -53,8 +54,9 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 65
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 75
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -82,7 +84,7 @@ steps:
   
 - label: Spec Decode Ngram + Suffix
   key: spec-decode-ngram-suffix
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
   device: h200_18gb
   source_file_dependencies:
     - vllm/v1/spec_decode/
@@ -92,10 +94,9 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 65
-      # TODO(akaratza): Test after Torch >= 2.12 bump
-      soft_fail: true
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 35
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -109,7 +110,7 @@ steps:
 
 - label: Spec Decode Draft Model
   key: spec-decode-draft-model
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   device: h200_18gb
   source_file_dependencies:
     - vllm/v1/spec_decode/
@@ -119,8 +120,9 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 50
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -134,7 +136,7 @@ steps:
 
 - label: Spec Decode Draft Model Nightly B200
   key: spec-decode-draft-model-nightly-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
   device: b200-k8s
   optional: true
   source_file_dependencies:
@@ -146,7 +148,7 @@ steps:
 
 - label: Speculators Correctness
   key: speculators-correctness
-  timeout_in_minutes: 60
+  timeout_in_minutes: 30
   device: h100
   optional: true
   num_devices: 1
@@ -159,7 +161,7 @@ steps:
     - pytest -v -s v1/spec_decode/test_speculators_correctness.py -m slow_test
 
 - label: Spec Decode MTP hybrid (B200)
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
   device: b200-k8s
   optional: true
   source_file_dependencies:
@@ -170,3 +172,19 @@ steps:
     - tests/v1/e2e/spec_decode/
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "qwen3_5-hybrid"
+
+- label: Spec Decode DeepSeek MTP Parallel Load (B200)
+  key: spec-decode-deepseek-mtp-parallel-load-b200
+  timeout_in_minutes: 30
+  device: b200-k8s
+  optional: true
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/v1/spec_decode/llm_base_proposer.py
+    - vllm/v1/spec_decode/eagle.py
+    - vllm/v1/worker/gpu/spec_decode/eagle/
+    - vllm/model_executor/models/deepseek_mtp.py
+    - vllm/model_executor/models/deepseek_v2.py
+    - tests/v1/e2e/spec_decode/test_mtp_parallel_load.py
+  commands:
+    - pytest -v -s v1/e2e/spec_decode/test_mtp_parallel_load.py

.buildkite/test_areas/weight_loading.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Weight Loading Multiple GPU  # 33min
   key: weight-loading-multiple-gpu
-  timeout_in_minutes: 45
+  timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   optional: true
@@ -15,7 +15,9 @@ steps:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
   mirror:
     amd:
+      dind: false
       device: mi300_2
+      timeout_in_minutes: 35
       depends_on:
       - image-build-amd
       commands:

.dockerignore

@@ -3,6 +3,7 @@
 dist
 vllm/*.so
 vllm/vllm-rs
+.git
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

.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
@@ -172,7 +173,7 @@ mkdocs.yaml @hmellor
 # Kernels
 /vllm/v1/attention/ops/chunked_prefill_paged_decode.py @tdoublep
 /vllm/v1/attention/ops/triton_unified_attention.py @tdoublep
-/vllm/model_executor/layers/fla @ZJY0516 @vadiklyutiy
+/vllm/third_party/flash_linear_attention @ZJY0516 @vadiklyutiy
 
 # ROCm related: specify owner with write access to notify AMD folks for careful code review
 /vllm/**/*rocm* @tjtanaa @dllehr-amd

.github/mergify.yml

@@ -181,6 +181,18 @@ pull_request_rules:
       add:
         - performance
 
+- name: label-quantization
+  description: Automatically apply quantization label
+  conditions:
+    - label != stale
+    - or:
+      - files~=^vllm/model_executor/layers/quantization/
+      - title~=(?i)quant
+  actions:
+    label:
+      add:
+        - quantization
+
 - name: label-qwen
   description: Automatically apply qwen label
   conditions:

.github/workflows/issue_autolabel.yml

@@ -130,6 +130,47 @@ jobs:
                   },
                 ],
               },
+              quantization: {
+                keywords: [
+                  {
+                    term: "quantization",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "quantized",
+                    searchIn: "both"
+                  },
+                ],
+              },
+              "intel-gpu": {
+                // Keyword search - matches whole words only (with word boundaries)
+                keywords: [
+                  {
+                    term: "B50",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "B60",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "B70",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "intel gpu",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "Arc GPU",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "BMG",
+                    searchIn: "both"
+                  },
+                ],
+              },
               // Add more label configurations here as needed
               // example: {
               //   keywords: [...],
@@ -323,7 +364,7 @@ jobs:
             // {users} will be replaced with @mentions
             const ccConfig = {
               rocm: {
-                users: ['hongxiayang', 'tjtanaa', 'vllmellm'],
+                users: ['hongxiayang', 'tjtanaa', 'vllmellm', 'giuseppegrossi'],
                 message: 'CC {users} for ROCm-related issue',
               },
               mistral: {
@@ -491,4 +532,4 @@ jobs:
               issue_number: context.issue.number,
               body: message,
             });
-            core.notice(`Requested missing ROCm info from @${author}: ${missing.map(m => m.name).join(', ')}`);
+            core.notice(`Requested missing ROCm info from @${author}: ${missing.map(m => m.name).join(', ')}`);

.gitignore

@@ -18,6 +18,9 @@ vllm/third_party/deep_gemm/
 # fmha_sm100 vendored package built from source
 vllm/third_party/fmha_sm100/
 
+# tml-fa4 vendored package built from source
+vllm/third_party/tml_fa4/
+
 # triton jit
 .triton
 

.pre-commit-config.yaml

@@ -4,7 +4,7 @@ default_install_hook_types:
 default_stages:
   - pre-commit # Run locally
   - manual # Run in CI
-exclude: 'vllm/third_party/.*'
+exclude: 'vllm/third_party/.*|vllm/models/kimi_k3/nvidia/ops/third_party/.*|vllm/models/kimi_k3/amd/ops/third_party/.*'
 repos:
 - repo: https://github.com/astral-sh/ruff-pre-commit
   rev: v0.14.0
@@ -30,7 +30,7 @@ repos:
   - id: markdownlint-cli2
     language_version: lts
     args: [--fix]
-    exclude: ^CLAUDE\.md$
+    exclude: (^|/)CLAUDE\.md$
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:
@@ -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

.yapfignore

@@ -1,2 +0,0 @@
-collect_env.py
-vllm/model_executor/layers/fla/ops/*.py

CMakeLists.txt

@@ -68,8 +68,17 @@ endif()
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.11.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.11.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.13.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.13.0")
+# TORCH_NIGHTLY=1 builds run against unpinned nightly wheels, so the supported-
+# version check would always warn. Only treat it as a nightly build when the
+# value is exactly "1" (the bootstrap exports TORCH_NIGHTLY=0 by default, which
+# must NOT suppress the warning for normal builds).
+if (DEFINED ENV{TORCH_NIGHTLY} AND "$ENV{TORCH_NIGHTLY}" STREQUAL "1")
+  set(TORCH_NIGHTLY_BUILD TRUE)
+else()
+  set(TORCH_NIGHTLY_BUILD FALSE)
+endif()
 
 #
 # Try to find python package with an executable that exactly matches
@@ -105,6 +114,11 @@ find_package(Torch REQUIRED)
 # Supported NVIDIA architectures.
 # This check must happen after find_package(Torch) because that's when CMAKE_CUDA_COMPILER_VERSION gets defined
 if(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
+   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.4)
+  # Rubin (10.7) can run SM100 family code, but CUDA 13.4 also supports
+  # targeting it directly.
+  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.7;11.0;12.0")
+elseif(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
    CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
    # starting from CUDA 12.9 and Blackwell (10.0), we use family-specific targets (10.0f, 12.0f, etc)
    # to support the whole generation without specifying all sub-architectures
@@ -175,7 +189,7 @@ endif()
 if (NOT HIP_FOUND AND NOT PYTORCH_FOUND_HIP AND CUDA_FOUND)
   set(VLLM_GPU_LANG "CUDA")
 
-  if (NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_CUDA})
+  if (NOT TORCH_NIGHTLY_BUILD AND NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_CUDA})
     message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_CUDA} "
       "expected for CUDA build, saw ${Torch_VERSION} instead.")
   endif()
@@ -188,7 +202,7 @@ elseif(HIP_FOUND OR PYTORCH_FOUND_HIP)
   enable_language(HIP)
 
   # ROCm 5.X and 6.X
-  if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
+  if (NOT TORCH_NIGHTLY_BUILD AND ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
       Torch_VERSION VERSION_LESS ${TORCH_SUPPORTED_VERSION_ROCM})
     message(WARNING "Pytorch version >= ${TORCH_SUPPORTED_VERSION_ROCM} "
       "expected for ROCm build, saw ${Torch_VERSION} instead.")
@@ -381,6 +395,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
     "csrc/libtorch_stable/cuda_view.cu"
     "csrc/libtorch_stable/cuda_utils_kernels.cu"
     "csrc/libtorch_stable/activation_kernels.cu"
+    "csrc/libtorch_stable/ngram_embedding_kernels.cu"
     "csrc/libtorch_stable/quantization/activation_kernels.cu"
     "csrc/libtorch_stable/quantization/w8a8/int8/scaled_quant.cu"
     "csrc/libtorch_stable/quantization/w8a8/fp8/common.cu"
@@ -401,8 +416,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
     "csrc/libtorch_stable/mamba/selective_scan_fwd.cu"
     "csrc/libtorch_stable/cache_kernels.cu"
     "csrc/libtorch_stable/cache_kernels_fused.cu"
+    "csrc/libtorch_stable/custom_all_gather_reduce_scatter.cu"
+    "csrc/libtorch_stable/custom_all_gather_reduce_scatter_ops.cpp"
     "csrc/libtorch_stable/custom_all_reduce.cu"
-    "csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu")
+    "csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu"
+    "csrc/libtorch_stable/fused_kimi_k3_mla_key_concat_kv_cache_kernel.cu")
 
   if(VLLM_GPU_LANG STREQUAL "CUDA" AND
      DEFINED CMAKE_CUDA_COMPILER_VERSION AND
@@ -410,7 +428,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
 
     if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
       cuda_archs_loose_intersection(COOPERATIVE_TOPK_ARCHS
-        "9.0a;10.0f;10.1f;10.3f;11.0f;12.0f;12.1f" "${CUDA_ARCHS}")
+        "9.0a;10.0f;10.1f;10.3f;10.7f;11.0f;12.0f;12.1f" "${CUDA_ARCHS}")
     else()
       cuda_archs_loose_intersection(COOPERATIVE_TOPK_ARCHS
         "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
@@ -685,7 +703,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
 
   # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;10.7f;11.0f;12.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
   endif()
@@ -805,7 +823,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
   # require CUDA 12.8 or later
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
   endif()
@@ -889,7 +907,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   endif()
 
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
   endif()
@@ -914,7 +932,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
 
   # moe_data.cu is used by all CUTLASS MoE kernels.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;10.7f;11.0f;12.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
   endif()
@@ -971,7 +989,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   # SM10x/11x FP4 kernels. MXFP4 experts quantization is currently compiled
   # only in this block; SM12x has separate NVFP4 matmul/MoE kernels above.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
   endif()
@@ -1037,7 +1055,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   # Runtime dispatch is gated in
   # vllm/v1/attention/backends/mla/cutlass_mla.py.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(MLA_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(MLA_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(MLA_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
   endif()
@@ -1059,6 +1077,41 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
     set(MLA_ARCHS)
   endif()
 
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(FUSED_KDA_DECODE_ARCHS
+      "9.0a;10.0f;12.0f" "${CUDA_ARCHS}")
+  endif()
+  if(FUSED_KDA_DECODE_ARCHS)
+    set(FUSED_KDA_DECODE_SRC
+      "csrc/libtorch_stable/kimi_k3/fused_kda_decode_kernel.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${FUSED_KDA_DECODE_SRC}"
+      CUDA_ARCHS "${FUSED_KDA_DECODE_ARCHS}")
+    set_property(SOURCE ${FUSED_KDA_DECODE_SRC} APPEND PROPERTY
+      COMPILE_OPTIONS "$<$<COMPILE_LANGUAGE:CUDA>:--use_fast_math>")
+    list(APPEND VLLM_STABLE_EXT_SRC "${FUSED_KDA_DECODE_SRC}")
+    message(STATUS
+      "Building fused KDA decode for archs: ${FUSED_KDA_DECODE_ARCHS}")
+  endif()
+
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(KIMI_K3_ATTN_RES_ARCHS
+      "10.0f" "${CUDA_ARCHS}")
+  endif()
+  if(KIMI_K3_ATTN_RES_ARCHS)
+    set(KIMI_K3_ATTN_RES_SRC
+      "csrc/libtorch_stable/kimi_k3/attn_res_kernel.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${KIMI_K3_ATTN_RES_SRC}"
+      CUDA_ARCHS "${KIMI_K3_ATTN_RES_ARCHS}")
+    set_property(SOURCE ${KIMI_K3_ATTN_RES_SRC} APPEND PROPERTY
+      COMPILE_OPTIONS
+      "$<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr;--expt-extended-lambda;--use_fast_math>")
+    list(APPEND VLLM_STABLE_EXT_SRC "${KIMI_K3_ATTN_RES_SRC}")
+    message(STATUS
+      "Building Kimi K3 AttnRes for archs: ${KIMI_K3_ATTN_RES_ARCHS}")
+  endif()
+
   # Hadacore kernels
   cuda_archs_loose_intersection(HADACORE_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
   if(HADACORE_ARCHS)
@@ -1100,6 +1153,14 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
       target_compile_definitions(_C_stable_libtorch PRIVATE
         VLLM_ENABLE_COOPERATIVE_TOPK=1)
     endif()
+    if(FUSED_KDA_DECODE_ARCHS)
+      target_compile_definitions(_C_stable_libtorch PRIVATE
+        VLLM_ENABLE_FUSED_KDA_DECODE=1)
+    endif()
+    if(KIMI_K3_ATTN_RES_ARCHS)
+      target_compile_definitions(_C_stable_libtorch PRIVATE
+        VLLM_ENABLE_KIMI_K3_ATTN_RES=1)
+    endif()
     # Needed by CUTLASS kernels
     target_compile_definitions(_C_stable_libtorch PRIVATE
       CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
@@ -1354,6 +1415,7 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
   set(VLLM_ROCM_EXT_SRC
     "csrc/rocm/torch_bindings.cpp"
     "csrc/rocm/skinny_gemms.cu"
+    "csrc/rocm/skinny_gemms_int4.cu"
     "csrc/rocm/attention.cu")
 
   set(VLLM_ROCM_HAS_GFX1100 OFF)
@@ -1396,7 +1458,9 @@ if (VLLM_GPU_LANG STREQUAL "CUDA")
     include(cmake/external_projects/deepgemm.cmake)
     include(cmake/external_projects/fmha_sm100.cmake)
     include(cmake/external_projects/flashmla.cmake)
+    include(cmake/external_projects/flashkda.cmake)
     include(cmake/external_projects/qutlass.cmake)
+    include(cmake/external_projects/tml_fa4.cmake)
 
     # vllm-flash-attn should be last as it overwrites some CMake functions
     include(cmake/external_projects/vllm_flash_attn.cmake)

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/benchmark.py

@@ -75,7 +75,11 @@ def run_mla_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
     from mla_runner import run_mla_benchmark as run_mla
 
     return run_mla(
-        config.backend, config, prefill_backend=config.prefill_backend, **kwargs
+        config.backend,
+        config,
+        prefill_backend=config.prefill_backend,
+        sparse_mla_force_mqa=config.sparse_mla_force_mqa,
+        **kwargs,
     )
 
 
@@ -592,6 +596,30 @@ def main():
         default="profile",
         help="Output file name for ncu profile (default: 'profile').",
     )
+    parser.add_argument(
+        "--torch-profile",
+        action="store_true",
+        default=False,
+        help="Collect a PyTorch profiler Chrome trace for each benchmark run.",
+    )
+    parser.add_argument(
+        "--torch-profile-dir",
+        default=None,
+        help="Directory for PyTorch profiler traces.",
+    )
+    parser.add_argument(
+        "--torch-profile-iters",
+        type=int,
+        default=3,
+        help="Number of forward passes to record per PyTorch profiler trace.",
+    )
+    parser.add_argument(
+        "--sparse-mla-mha-variants",
+        nargs="+",
+        default=None,
+        choices=["dense_mha", "mqa"],
+        help="Sparse MLA variants to run in mha_vs_mqa mode. Defaults to both.",
+    )
 
     # Parameter sweep (use YAML config for advanced sweeps)
     parser.add_argument(
@@ -641,6 +669,7 @@ def main():
 
         # Prefill backends (e.g., ["fa3", "fa4"])
         args.prefill_backends = yaml_config.get("prefill_backends", None)
+        args.prefill_backend = yaml_config.get("prefill_backend", None)
 
         # FP8 output benchmark knobs; CLI wins.
         if args.fp8_output_scale is None:
@@ -683,6 +712,9 @@ def main():
             args.num_q_heads = model.get("num_q_heads", args.num_q_heads)
             args.num_kv_heads = model.get("num_kv_heads", args.num_kv_heads)
             args.block_size = model.get("block_size", args.block_size)
+            args.max_model_len = model.get(
+                "max_model_len", getattr(args, "max_model_len", None)
+            )
             # MLA-specific dimensions
             args.kv_lora_rank = model.get("kv_lora_rank", args.kv_lora_rank)
             args.qk_nope_head_dim = model.get("qk_nope_head_dim", args.qk_nope_head_dim)
@@ -701,6 +733,21 @@ def main():
             args.cuda_graphs = yaml_config["cuda_graphs"]
         if "ncu_profile" in yaml_config:
             args.ncu_profile = yaml_config["ncu_profile"]
+        if "torch_profile" in yaml_config:
+            args.torch_profile = yaml_config["torch_profile"]
+        if "torch_profile_dir" in yaml_config:
+            args.torch_profile_dir = yaml_config["torch_profile_dir"]
+        if "torch_profile_iters" in yaml_config:
+            args.torch_profile_iters = yaml_config["torch_profile_iters"]
+        args.sparse_mla_topk_pattern = yaml_config.get(
+            "sparse_mla_topk_pattern", "random"
+        )
+        args.sparse_mla_dense_mha_max_seq_len = yaml_config.get(
+            "sparse_mla_dense_mha_max_seq_len", None
+        )
+        args.sparse_mla_mha_variants = yaml_config.get(
+            "sparse_mla_mha_variants", args.sparse_mla_mha_variants
+        )
 
         # Parameter sweep configuration
         if "parameter_sweep" in yaml_config:
@@ -842,8 +889,6 @@ def main():
                         num_kv_heads=args.num_kv_heads,
                         block_size=args.block_size,
                         device=args.device,
-                        repeats=args.repeats,
-                        warmup_iters=args.warmup_iters,
                         profile_memory=args.profile_memory,
                         kv_cache_dtype=args.kv_cache_dtype,
                         use_cuda_graphs=args.cuda_graphs,
@@ -1063,6 +1108,133 @@ def main():
                     f"\n  [yellow]Prefill always faster for batch_size={bs}[/]"
                 )
 
+    # Handle MHA vs MQA comparison mode for sparse MLA
+    elif hasattr(args, "mode") and args.mode == "mha_vs_mqa":
+        console.print("[yellow]Mode: MHA vs MQA comparison for sparse MLA[/]")
+
+        sparse_mla_topk_pattern = getattr(args, "sparse_mla_topk_pattern", "random")
+        dense_mha_max_seq_len = getattr(args, "sparse_mla_dense_mha_max_seq_len", None)
+        prefill_backend = getattr(args, "prefill_backend", None)
+        if prefill_backend:
+            console.print(f"Prefill backend: {prefill_backend}")
+        available_variants = [
+            ("dense_mha", False, "dense"),
+            ("mqa", True, "auto"),
+        ]
+        requested_variants = getattr(args, "sparse_mla_mha_variants", None)
+        if requested_variants is not None:
+            valid_variants = {label for label, _, _ in available_variants}
+            invalid_variants = sorted(set(requested_variants) - valid_variants)
+            if invalid_variants:
+                raise ValueError(
+                    "Invalid sparse_mla_mha_variants entries: "
+                    f"{invalid_variants}. Valid variants are: "
+                    f"{sorted(valid_variants)}"
+                )
+            requested_variant_set = set(requested_variants)
+            variants = [
+                variant
+                for variant in available_variants
+                if variant[0] in requested_variant_set
+            ]
+        else:
+            variants = available_variants
+        formatter = ResultsFormatter(console)
+        total = 0
+        for spec in args.batch_specs:
+            q_len = max(request.q_len for request in parse_batch_spec(spec))
+            for variant_label, _, _ in variants:
+                if (
+                    variant_label == "dense_mha"
+                    and dense_mha_max_seq_len is not None
+                    and q_len > dense_mha_max_seq_len
+                ):
+                    continue
+                total += len(backends)
+
+        with tqdm(total=total, desc="Benchmarking") as pbar:
+            for spec in args.batch_specs:
+                q_len = max(request.q_len for request in parse_batch_spec(spec))
+                for backend in backends:
+                    for variant_label, force_mqa, mha_mode in variants:
+                        if (
+                            variant_label == "dense_mha"
+                            and dense_mha_max_seq_len is not None
+                            and q_len > dense_mha_max_seq_len
+                        ):
+                            continue
+                        config = BenchmarkConfig(
+                            backend=f"{backend}_{variant_label}",
+                            batch_spec=spec,
+                            num_layers=args.num_layers,
+                            head_dim=args.head_dim,
+                            num_q_heads=args.num_q_heads,
+                            num_kv_heads=args.num_kv_heads,
+                            block_size=args.block_size,
+                            device=args.device,
+                            max_model_len=getattr(args, "max_model_len", None),
+                            kv_cache_dtype=args.kv_cache_dtype,
+                            profile_memory=args.profile_memory,
+                            use_cuda_graphs=args.cuda_graphs,
+                            ncu_profile=args.ncu_profile,
+                            torch_profile=args.torch_profile,
+                            torch_profile_dir=args.torch_profile_dir,
+                            torch_profile_iters=args.torch_profile_iters,
+                            warmup_ms=args.warmup_ms,
+                            kv_lora_rank=getattr(args, "kv_lora_rank", None),
+                            qk_nope_head_dim=getattr(args, "qk_nope_head_dim", None),
+                            qk_rope_head_dim=getattr(args, "qk_rope_head_dim", None),
+                            v_head_dim=getattr(args, "v_head_dim", None),
+                            sparse_mla_force_mqa=force_mqa,
+                            sparse_mla_mha_mode=mha_mode,
+                            sparse_mla_dense_mha_max_seq_len=dense_mha_max_seq_len,
+                            sparse_mla_topk_pattern=sparse_mla_topk_pattern,
+                            prefill_backend=prefill_backend,
+                        )
+
+                        # run_mla_benchmark needs the real backend name
+                        from mla_runner import run_mla_benchmark as run_mla
+
+                        run_label = f"{backend}_{variant_label} {spec}"
+                        pbar.set_postfix_str(run_label)
+
+                        try:
+                            result = run_mla(
+                                backend,
+                                config,
+                                prefill_backend=prefill_backend,
+                                sparse_mla_force_mqa=force_mqa,
+                            )
+                        except Exception as e:
+                            result = BenchmarkResult(
+                                config=config,
+                                mean_time=float("inf"),
+                                median_time=float("inf"),
+                                std_time=0,
+                                min_time=float("inf"),
+                                max_time=float("inf"),
+                                error=str(e),
+                            )
+
+                        all_results.append(result)
+                        if args.output_csv:
+                            formatter.save_csv(all_results, args.output_csv)
+                        if args.output_json:
+                            formatter.save_json(all_results, args.output_json)
+
+                        if not result.success:
+                            console.print(
+                                f"[red]Error {backend}_{variant_label} "
+                                f"{spec}: {result.error}[/]"
+                            )
+
+                        pbar.update(1)
+
+        # Display results with variant labels as separate "backends"
+        console.print("\n[bold green]MHA vs MQA Results:[/]")
+        variant_backends = [f"{b}_{v}" for b in backends for v, _, _ in variants]
+        formatter.print_table(all_results, variant_backends)
+
     # Handle model parameter sweep mode
     elif hasattr(args, "model_parameter_sweep") and args.model_parameter_sweep:
         # Model parameter sweep

benchmarks/attention_benchmarks/common.py

@@ -4,8 +4,10 @@
 """Common utilities for attention benchmarking."""
 
 import csv
+import gc
 import json
 import math
+from collections.abc import Sequence
 from dataclasses import asdict, dataclass
 from pathlib import Path
 from typing import Any
@@ -44,10 +46,13 @@ def run_do_bench(
     kwargs: dict[str, Any] = {"return_mode": "all"}
     if use_cuda_graphs:
         result = triton.testing.do_bench_cudagraph(benchmark_fn, **kwargs)
+        gc.collect()
+        torch.accelerator.empty_cache()
     else:
         if warmup_ms is not None:
             kwargs["warmup"] = warmup_ms
         result = triton.testing.do_bench(benchmark_fn, **kwargs)
+    torch.accelerator.synchronize()
     return result
 
 
@@ -91,42 +96,6 @@ except ImportError:
     AttentionLayerBase = object  # Fallback
 
 
-class MockKVBProj:
-    """Mock KV projection layer for MLA prefill mode.
-
-    Mimics ColumnParallelLinear behavior for kv_b_proj in MLA backends.
-    Projects kv_c_normed to [qk_nope_head_dim + v_head_dim] per head.
-    """
-
-    def __init__(self, num_heads: int, qk_nope_head_dim: int, v_head_dim: int):
-        self.num_heads = num_heads
-        self.qk_nope_head_dim = qk_nope_head_dim
-        self.v_head_dim = v_head_dim
-        self.out_dim = qk_nope_head_dim + v_head_dim
-        self.weight = torch.empty(0, dtype=torch.bfloat16)
-
-    def __call__(self, x: torch.Tensor) -> tuple[torch.Tensor]:
-        """
-        Project kv_c_normed to output space.
-
-        Args:
-            x: Input tensor [num_tokens, kv_lora_rank]
-
-        Returns:
-            Tuple containing output tensor
-                [num_tokens, num_heads, qk_nope_head_dim + v_head_dim]
-        """
-        num_tokens = x.shape[0]
-        result = torch.randn(
-            num_tokens,
-            self.num_heads,
-            self.out_dim,
-            device=x.device,
-            dtype=x.dtype,
-        )
-        return (result,)  # Return as tuple to match ColumnParallelLinear API
-
-
 class MockIndexer:
     """Mock Indexer for sparse MLA backends.
 
@@ -158,6 +127,60 @@ class MockIndexer:
         )
         self.topk_indices_buffer[:num_tokens] = indices
 
+    def fill_indices(
+        self,
+        num_tokens: int,
+        max_kv_len: int,
+        pattern: str = "random",
+        requests: Sequence[Any] | None = None,
+    ):
+        if pattern == "random":
+            self.fill_random_indices(num_tokens, max_kv_len)
+            return
+        if pattern == "prefix":
+            indices = torch.arange(
+                self.topk_tokens,
+                dtype=torch.int32,
+                device=self.topk_indices_buffer.device,
+            )
+            indices = (indices % max_kv_len).expand(num_tokens, -1)
+            self.topk_indices_buffer[:num_tokens] = indices
+            return
+        if pattern == "sliding_window":
+            if requests is None:
+                start = max(max_kv_len - self.topk_tokens, 0)
+                indices = torch.arange(
+                    start,
+                    start + self.topk_tokens,
+                    dtype=torch.int32,
+                    device=self.topk_indices_buffer.device,
+                )
+                indices = indices.clamp(max=max_kv_len - 1).expand(num_tokens, -1)
+                self.topk_indices_buffer[:num_tokens] = indices
+                return
+
+            rows = []
+            offsets = torch.arange(
+                self.topk_tokens,
+                dtype=torch.int32,
+                device=self.topk_indices_buffer.device,
+            ) - (self.topk_tokens - 1)
+            for request in requests:
+                q_len = request.q_len
+                kv_len = request.kv_len
+                context_len = kv_len - q_len
+                positions = torch.arange(
+                    context_len,
+                    kv_len,
+                    dtype=torch.int32,
+                    device=self.topk_indices_buffer.device,
+                )
+                row_indices = positions[:, None] + offsets[None, :]
+                rows.append(row_indices.clamp(min=0, max=kv_len - 1))
+            self.topk_indices_buffer[:num_tokens] = torch.cat(rows, dim=0)
+            return
+        raise ValueError(f"Unknown sparse MLA topk pattern: {pattern}")
+
 
 class MockLayer(AttentionLayerBase):
     """Mock attention layer with scale parameters and impl.
@@ -252,10 +275,14 @@ class BenchmarkConfig:
     num_kv_heads: int
     block_size: int
     device: str
+    max_model_len: int | None = None
     dtype: torch.dtype = torch.float16
     profile_memory: bool = False
-    use_cuda_graphs: bool = False
+    use_cuda_graphs: bool = True
     ncu_profile: bool = False
+    torch_profile: bool = False
+    torch_profile_dir: str | None = None
+    torch_profile_iters: int = 3
     warmup_ms: int | None = None
 
     # "auto" or "fp8"
@@ -271,6 +298,10 @@ class BenchmarkConfig:
     # Backend-specific tuning
     num_kv_splits: int | None = None  # CUTLASS MLA
     reorder_batch_threshold: int | None = None  # FlashAttn MLA, FlashMLA
+    sparse_mla_force_mqa: bool = False  # Force MQA path for sparse MLA
+    sparse_mla_mha_mode: str = "auto"  # "auto" or "dense"
+    sparse_mla_dense_mha_max_seq_len: int | None = None
+    sparse_mla_topk_pattern: str = "random"  # "random", "prefix", "sliding_window"
     num_splits: int | None = None  # FlashAttention split-K (0=auto, 1=disabled)
 
 

benchmarks/attention_benchmarks/configs/mla_sparse_mha_vs_mqa.yaml

@@ -0,0 +1,474 @@
+# Sparse MLA benchmark: forward_mha vs forward_mqa
+#
+# Usage:
+#   python benchmark.py --config configs/mla_sparse_mha_vs_mqa.yaml
+#
+# Heatmap grid:
+#   - batch_size: 1, 2, 4, 8, 16, 32
+#   - seq_len: 32, 64, 128, 256, 512, 1024, 2048
+#   - q_len: powers of two through seq_len
+#
+# Specs with q_len < seq_len include context; the q_len == seq_len diagonal
+# covers pure prefill.
+# The model shape below is the DP case. For the TP8 run, manually change
+# model.num_q_heads from 128 to 16 before rerunning this benchmark.
+
+mode: mha_vs_mqa
+
+model:
+  name: "deepseek-v3"
+  num_layers: 60
+  num_q_heads: 128
+  num_kv_heads: 1
+  head_dim: 576
+  kv_lora_rank: 512
+  qk_nope_head_dim: 128
+  qk_rope_head_dim: 64
+  v_head_dim: 128
+  block_size: 128
+  max_model_len: 2048
+
+batch_specs:
+  # Batch size 1
+  # seq_len = 32
+  - "1q1s32"
+  - "1q2s32"
+  - "1q4s32"
+  - "1q8s32"
+  - "1q16s32"
+  - "1q32"
+  # seq_len = 64
+  - "1q1s64"
+  - "1q2s64"
+  - "1q4s64"
+  - "1q8s64"
+  - "1q16s64"
+  - "1q32s64"
+  - "1q64"
+  # seq_len = 128
+  - "1q1s128"
+  - "1q2s128"
+  - "1q4s128"
+  - "1q8s128"
+  - "1q16s128"
+  - "1q32s128"
+  - "1q64s128"
+  - "1q128"
+  # seq_len = 256
+  - "1q1s256"
+  - "1q2s256"
+  - "1q4s256"
+  - "1q8s256"
+  - "1q16s256"
+  - "1q32s256"
+  - "1q64s256"
+  - "1q128s256"
+  - "1q256"
+  # seq_len = 512
+  - "1q1s512"
+  - "1q2s512"
+  - "1q4s512"
+  - "1q8s512"
+  - "1q16s512"
+  - "1q32s512"
+  - "1q64s512"
+  - "1q128s512"
+  - "1q256s512"
+  - "1q512"
+  # seq_len = 1024
+  - "1q1s1024"
+  - "1q2s1024"
+  - "1q4s1024"
+  - "1q8s1024"
+  - "1q16s1024"
+  - "1q32s1024"
+  - "1q64s1024"
+  - "1q128s1024"
+  - "1q256s1024"
+  - "1q512s1024"
+  - "1q1024"
+  # seq_len = 2048
+  - "1q1s2048"
+  - "1q2s2048"
+  - "1q4s2048"
+  - "1q8s2048"
+  - "1q16s2048"
+  - "1q32s2048"
+  - "1q64s2048"
+  - "1q128s2048"
+  - "1q256s2048"
+  - "1q512s2048"
+  - "1q1024s2048"
+  - "1q2048"
+
+  # Batch size 2
+  # seq_len = 32
+  - "2q1s32"
+  - "2q2s32"
+  - "2q4s32"
+  - "2q8s32"
+  - "2q16s32"
+  - "2q32"
+  # seq_len = 64
+  - "2q1s64"
+  - "2q2s64"
+  - "2q4s64"
+  - "2q8s64"
+  - "2q16s64"
+  - "2q32s64"
+  - "2q64"
+  # seq_len = 128
+  - "2q1s128"
+  - "2q2s128"
+  - "2q4s128"
+  - "2q8s128"
+  - "2q16s128"
+  - "2q32s128"
+  - "2q64s128"
+  - "2q128"
+  # seq_len = 256
+  - "2q1s256"
+  - "2q2s256"
+  - "2q4s256"
+  - "2q8s256"
+  - "2q16s256"
+  - "2q32s256"
+  - "2q64s256"
+  - "2q128s256"
+  - "2q256"
+  # seq_len = 512
+  - "2q1s512"
+  - "2q2s512"
+  - "2q4s512"
+  - "2q8s512"
+  - "2q16s512"
+  - "2q32s512"
+  - "2q64s512"
+  - "2q128s512"
+  - "2q256s512"
+  - "2q512"
+  # seq_len = 1024
+  - "2q1s1024"
+  - "2q2s1024"
+  - "2q4s1024"
+  - "2q8s1024"
+  - "2q16s1024"
+  - "2q32s1024"
+  - "2q64s1024"
+  - "2q128s1024"
+  - "2q256s1024"
+  - "2q512s1024"
+  - "2q1024"
+  # seq_len = 2048
+  - "2q1s2048"
+  - "2q2s2048"
+  - "2q4s2048"
+  - "2q8s2048"
+  - "2q16s2048"
+  - "2q32s2048"
+  - "2q64s2048"
+  - "2q128s2048"
+  - "2q256s2048"
+  - "2q512s2048"
+  - "2q1024s2048"
+  - "2q2048"
+
+  # Batch size 4
+  # seq_len = 32
+  - "4q1s32"
+  - "4q2s32"
+  - "4q4s32"
+  - "4q8s32"
+  - "4q16s32"
+  - "4q32"
+  # seq_len = 64
+  - "4q1s64"
+  - "4q2s64"
+  - "4q4s64"
+  - "4q8s64"
+  - "4q16s64"
+  - "4q32s64"
+  - "4q64"
+  # seq_len = 128
+  - "4q1s128"
+  - "4q2s128"
+  - "4q4s128"
+  - "4q8s128"
+  - "4q16s128"
+  - "4q32s128"
+  - "4q64s128"
+  - "4q128"
+  # seq_len = 256
+  - "4q1s256"
+  - "4q2s256"
+  - "4q4s256"
+  - "4q8s256"
+  - "4q16s256"
+  - "4q32s256"
+  - "4q64s256"
+  - "4q128s256"
+  - "4q256"
+  # seq_len = 512
+  - "4q1s512"
+  - "4q2s512"
+  - "4q4s512"
+  - "4q8s512"
+  - "4q16s512"
+  - "4q32s512"
+  - "4q64s512"
+  - "4q128s512"
+  - "4q256s512"
+  - "4q512"
+  # seq_len = 1024
+  - "4q1s1024"
+  - "4q2s1024"
+  - "4q4s1024"
+  - "4q8s1024"
+  - "4q16s1024"
+  - "4q32s1024"
+  - "4q64s1024"
+  - "4q128s1024"
+  - "4q256s1024"
+  - "4q512s1024"
+  - "4q1024"
+  # seq_len = 2048
+  - "4q1s2048"
+  - "4q2s2048"
+  - "4q4s2048"
+  - "4q8s2048"
+  - "4q16s2048"
+  - "4q32s2048"
+  - "4q64s2048"
+  - "4q128s2048"
+  - "4q256s2048"
+  - "4q512s2048"
+  - "4q1024s2048"
+  - "4q2048"
+
+  # Batch size 8
+  # seq_len = 32
+  - "8q1s32"
+  - "8q2s32"
+  - "8q4s32"
+  - "8q8s32"
+  - "8q16s32"
+  - "8q32"
+  # seq_len = 64
+  - "8q1s64"
+  - "8q2s64"
+  - "8q4s64"
+  - "8q8s64"
+  - "8q16s64"
+  - "8q32s64"
+  - "8q64"
+  # seq_len = 128
+  - "8q1s128"
+  - "8q2s128"
+  - "8q4s128"
+  - "8q8s128"
+  - "8q16s128"
+  - "8q32s128"
+  - "8q64s128"
+  - "8q128"
+  # seq_len = 256
+  - "8q1s256"
+  - "8q2s256"
+  - "8q4s256"
+  - "8q8s256"
+  - "8q16s256"
+  - "8q32s256"
+  - "8q64s256"
+  - "8q128s256"
+  - "8q256"
+  # seq_len = 512
+  - "8q1s512"
+  - "8q2s512"
+  - "8q4s512"
+  - "8q8s512"
+  - "8q16s512"
+  - "8q32s512"
+  - "8q64s512"
+  - "8q128s512"
+  - "8q256s512"
+  - "8q512"
+  # seq_len = 1024
+  - "8q1s1024"
+  - "8q2s1024"
+  - "8q4s1024"
+  - "8q8s1024"
+  - "8q16s1024"
+  - "8q32s1024"
+  - "8q64s1024"
+  - "8q128s1024"
+  - "8q256s1024"
+  - "8q512s1024"
+  - "8q1024"
+  # seq_len = 2048
+  - "8q1s2048"
+  - "8q2s2048"
+  - "8q4s2048"
+  - "8q8s2048"
+  - "8q16s2048"
+  - "8q32s2048"
+  - "8q64s2048"
+  - "8q128s2048"
+  - "8q256s2048"
+  - "8q512s2048"
+  - "8q1024s2048"
+  - "8q2048"
+
+  # Batch size 16
+  # seq_len = 32
+  - "16q1s32"
+  - "16q2s32"
+  - "16q4s32"
+  - "16q8s32"
+  - "16q16s32"
+  - "16q32"
+  # seq_len = 64
+  - "16q1s64"
+  - "16q2s64"
+  - "16q4s64"
+  - "16q8s64"
+  - "16q16s64"
+  - "16q32s64"
+  - "16q64"
+  # seq_len = 128
+  - "16q1s128"
+  - "16q2s128"
+  - "16q4s128"
+  - "16q8s128"
+  - "16q16s128"
+  - "16q32s128"
+  - "16q64s128"
+  - "16q128"
+  # seq_len = 256
+  - "16q1s256"
+  - "16q2s256"
+  - "16q4s256"
+  - "16q8s256"
+  - "16q16s256"
+  - "16q32s256"
+  - "16q64s256"
+  - "16q128s256"
+  - "16q256"
+  # seq_len = 512
+  - "16q1s512"
+  - "16q2s512"
+  - "16q4s512"
+  - "16q8s512"
+  - "16q16s512"
+  - "16q32s512"
+  - "16q64s512"
+  - "16q128s512"
+  - "16q256s512"
+  - "16q512"
+  # seq_len = 1024
+  - "16q1s1024"
+  - "16q2s1024"
+  - "16q4s1024"
+  - "16q8s1024"
+  - "16q16s1024"
+  - "16q32s1024"
+  - "16q64s1024"
+  - "16q128s1024"
+  - "16q256s1024"
+  - "16q512s1024"
+  - "16q1024"
+  # seq_len = 2048
+  - "16q1s2048"
+  - "16q2s2048"
+  - "16q4s2048"
+  - "16q8s2048"
+  - "16q16s2048"
+  - "16q32s2048"
+  - "16q64s2048"
+  - "16q128s2048"
+  - "16q256s2048"
+  - "16q512s2048"
+  - "16q1024s2048"
+  - "16q2048"
+
+  # Batch size 32
+  # seq_len = 32
+  - "32q1s32"
+  - "32q2s32"
+  - "32q4s32"
+  - "32q8s32"
+  - "32q16s32"
+  - "32q32"
+  # seq_len = 64
+  - "32q1s64"
+  - "32q2s64"
+  - "32q4s64"
+  - "32q8s64"
+  - "32q16s64"
+  - "32q32s64"
+  - "32q64"
+  # seq_len = 128
+  - "32q1s128"
+  - "32q2s128"
+  - "32q4s128"
+  - "32q8s128"
+  - "32q16s128"
+  - "32q32s128"
+  - "32q64s128"
+  - "32q128"
+  # seq_len = 256
+  - "32q1s256"
+  - "32q2s256"
+  - "32q4s256"
+  - "32q8s256"
+  - "32q16s256"
+  - "32q32s256"
+  - "32q64s256"
+  - "32q128s256"
+  - "32q256"
+  # seq_len = 512
+  - "32q1s512"
+  - "32q2s512"
+  - "32q4s512"
+  - "32q8s512"
+  - "32q16s512"
+  - "32q32s512"
+  - "32q64s512"
+  - "32q128s512"
+  - "32q256s512"
+  - "32q512"
+  # seq_len = 1024
+  - "32q1s1024"
+  - "32q2s1024"
+  - "32q4s1024"
+  - "32q8s1024"
+  - "32q16s1024"
+  - "32q32s1024"
+  - "32q64s1024"
+  - "32q128s1024"
+  - "32q256s1024"
+  - "32q512s1024"
+  - "32q1024"
+  # seq_len = 2048
+  - "32q1s2048"
+  - "32q2s2048"
+  - "32q4s2048"
+  - "32q8s2048"
+  - "32q16s2048"
+  - "32q32s2048"
+  - "32q64s2048"
+  - "32q128s2048"
+  - "32q256s2048"
+  - "32q512s2048"
+  - "32q1024s2048"
+  - "32q2048"
+
+backends:
+  - FLASHMLA_SPARSE
+
+device: "cuda:0"
+profile_memory: false
+sparse_mla_dense_mha_max_seq_len: 2048
+sparse_mla_topk_pattern: "random"
+
+output:
+  csv: "benchmark_output/mla_sparse_mha_vs_mqa.csv"
+  json: "benchmark_output/mla_sparse_mha_vs_mqa.json"

benchmarks/attention_benchmarks/mla_runner.py

@@ -9,6 +9,8 @@ needing full VllmConfig integration.
 """
 
 import statistics
+import tempfile
+from pathlib import Path
 
 import numpy as np
 import torch
@@ -17,7 +19,6 @@ from common import (
     BenchmarkResult,
     MockHfConfig,
     MockIndexer,
-    MockKVBProj,
     MockLayer,
     run_do_bench,
     run_ncu_profile,
@@ -33,8 +34,59 @@ from vllm.config import (
     VllmConfig,
     set_current_vllm_config,
 )
+from vllm.model_executor.layers.linear import ColumnParallelLinear
 from vllm.v1.attention.backends.mla.prefill.registry import MLAPrefillBackendEnum
 
+
+def _safe_profile_name(value: str) -> str:
+    return "".join(c if c.isalnum() or c in "._-" else "_" for c in value)
+
+
+def _create_kv_b_proj(
+    mla_dims: dict,
+    device: torch.device,
+):
+    kv_b_proj = ColumnParallelLinear(
+        mla_dims["kv_lora_rank"],
+        mla_dims["num_q_heads"]
+        * (mla_dims["qk_nope_head_dim"] + mla_dims["v_head_dim"]),
+        bias=False,
+        params_dtype=torch.bfloat16,
+        quant_config=None,
+        prefix="benchmark.kv_b_proj",
+    ).to(device)
+    with torch.no_grad():
+        kv_b_proj.weight.copy_(torch.randn_like(kv_b_proj.weight))
+    return kv_b_proj
+
+
+def _ensure_single_rank_model_parallel() -> None:
+    import torch.distributed as dist
+
+    from vllm.distributed import (
+        ensure_model_parallel_initialized,
+        init_distributed_environment,
+        model_parallel_is_initialized,
+    )
+
+    if not dist.is_available():
+        return
+    if not dist.is_initialized():
+        with tempfile.NamedTemporaryFile(
+            prefix="vllm_bench_dist_", delete=False
+        ) as init_file:
+            distributed_init_method = f"file://{init_file.name}"
+        init_distributed_environment(
+            world_size=1,
+            rank=0,
+            distributed_init_method=distributed_init_method,
+            local_rank=0,
+            backend="nccl",
+        )
+    if not model_parallel_is_initialized():
+        ensure_model_parallel_initialized(1, 1)
+
+
 # ============================================================================
 # VllmConfig Creation
 # ============================================================================
@@ -66,10 +118,12 @@ def create_minimal_vllm_config(
     block_size: int = 128,
     max_num_seqs: int = 256,
     max_num_batched_tokens: int = 8192,
+    max_model_len: int = 32768,
     mla_dims: dict | None = None,
     index_topk: int | None = None,
     prefill_backend: str | None = None,
     kv_cache_dtype: str = "auto",
+    sparse_mla_force_mqa: bool = False,
 ) -> VllmConfig:
     """
     Create minimal VllmConfig for MLA benchmarks.
@@ -86,6 +140,8 @@ def create_minimal_vllm_config(
         prefill_backend: Prefill backend name (e.g., "fa3", "fa4", "flashinfer",
                         "trtllm"). Configures the attention config to force
                         the specified prefill backend.
+        sparse_mla_force_mqa: If True, forces all sparse MLA tokens through
+                    forward_mqa (even prefill tokens).
 
     Returns:
         VllmConfig for benchmarking
@@ -131,7 +187,7 @@ def create_minimal_vllm_config(
             trust_remote_code=True,
             dtype="bfloat16",
             seed=0,
-            max_model_len=32768,
+            max_model_len=max_model_len,
             quantization=None,
             enforce_eager=False,
             max_logprobs=20,
@@ -163,7 +219,7 @@ def create_minimal_vllm_config(
     scheduler_config = SchedulerConfig(
         max_num_seqs=max_num_seqs,
         max_num_batched_tokens=max(max_num_batched_tokens, max_num_seqs),
-        max_model_len=32768,
+        max_model_len=max_model_len,
         is_encoder_decoder=False,
         enable_chunked_prefill=True,
     )
@@ -192,6 +248,9 @@ def create_minimal_vllm_config(
                 "flash_attn_version"
             ]
 
+    if sparse_mla_force_mqa:
+        vllm_config.attention_config.sparse_mla_force_mqa = True
+
     return vllm_config
 
 
@@ -548,12 +607,7 @@ def _create_backend_impl(
     # Calculate scale
     scale = 1.0 / np.sqrt(mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"])
 
-    # Create mock kv_b_proj layer for prefill mode
-    mock_kv_b_proj = MockKVBProj(
-        num_heads=mla_dims["num_q_heads"],
-        qk_nope_head_dim=mla_dims["qk_nope_head_dim"],
-        v_head_dim=mla_dims["v_head_dim"],
-    )
+    kv_b_proj = _create_kv_b_proj(mla_dims, device)
 
     # Create indexer for sparse backends
     indexer = None
@@ -584,7 +638,7 @@ def _create_backend_impl(
         "qk_rope_head_dim": mla_dims["qk_rope_head_dim"],
         "qk_head_dim": mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"],
         "v_head_dim": mla_dims["v_head_dim"],
-        "kv_b_proj": mock_kv_b_proj,
+        "kv_b_proj": kv_b_proj,
     }
 
     # Add indexer for sparse backends
@@ -785,14 +839,35 @@ def _run_single_benchmark(
     # Fill indexer with random indices for sparse backends
     is_sparse = backend_cfg.get("is_sparse", False)
     if is_sparse and indexer is not None:
-        indexer.fill_random_indices(total_q, max_kv_len)
+        indexer.fill_indices(
+            total_q,
+            max_kv_len,
+            getattr(config, "sparse_mla_topk_pattern", "random"),
+        )
 
     # Determine which forward methods to use based on metadata.
-    # Sparse MLA backends always use forward_mqa
-    has_decode = is_sparse or getattr(metadata, "decode", None) is not None
-    has_prefill = not is_sparse and getattr(metadata, "prefill", None) is not None
+    # Non-sparse backends use .decode/.prefill sub-objects.
+    # Sparse backends use num_decode_tokens/num_prefills directly.
+    #
+    # sparse_mla_force_mqa overrides: even for prefill metadata, use MQA.
+    force_mqa = getattr(config, "sparse_mla_force_mqa", False)
+    force_dense_mha = getattr(config, "sparse_mla_mha_mode", "auto") == "dense"
+    if force_mqa:
+        has_decode = True
+        has_prefill = False
+    elif is_sparse:
+        has_decode = metadata.num_decode_tokens > 0
+        has_prefill = metadata.num_prefills > 0
+    else:
+        has_decode = metadata.decode is not None
+        has_prefill = metadata.prefill is not None
     if not has_decode and not has_prefill:
         raise RuntimeError("Metadata has neither decode nor prefill metadata")
+    if is_sparse and force_dense_mha and not has_prefill:
+        raise RuntimeError(
+            "Sparse MLA dense_mha benchmark did not produce prefill metadata. "
+            "Check reorder_batch_threshold/path forcing."
+        )
 
     num_decode = (
         metadata.num_decode_tokens
@@ -871,7 +946,6 @@ def _run_single_benchmark(
                 metadata,
                 prefill_inputs["k_scale"],
                 prefill_fp8_output if fused_output else prefill_inputs["output"],
-                prefill_output_scale if fused_output else None,
             )
             if fused_output:
                 out = prefill_fp8_output
@@ -898,6 +972,48 @@ def _run_single_benchmark(
             throughput_tokens_per_sec=0.0,
         )
 
+    if config.torch_profile:
+        profile_dir = Path(
+            config.torch_profile_dir or "benchmark_outputs/torch_profiles"
+        )
+        profile_dir.mkdir(parents=True, exist_ok=True)
+        trace_name = _safe_profile_name(f"{config.backend}_{config.batch_spec}")
+        trace_path = profile_dir / f"{trace_name}.json"
+        iters = max(config.torch_profile_iters, 1)
+
+        forward_fn()
+        torch.accelerator.synchronize()
+        with torch.profiler.profile(
+            activities=[
+                torch.profiler.ProfilerActivity.CPU,
+                torch.profiler.ProfilerActivity.CUDA,
+            ],
+            record_shapes=True,
+            profile_memory=True,
+            with_stack=False,
+        ) as prof:
+            for _ in range(iters):
+                forward_fn()
+                torch.accelerator.synchronize()
+                prof.step()
+        prof.export_chrome_trace(str(trace_path))
+        print(f"Saved PyTorch profiler trace to {trace_path}")
+        print(
+            prof.key_averages().table(
+                sort_by="cuda_time_total",
+                row_limit=25,
+            )
+        )
+        return BenchmarkResult(
+            config=config,
+            mean_time=0.0,
+            median_time=0.0,
+            std_time=0.0,
+            min_time=0.0,
+            max_time=0.0,
+            throughput_tokens_per_sec=0.0,
+        )
+
     all_ms = run_do_bench(benchmark_fn, config.use_cuda_graphs, config.warmup_ms)
 
     # Convert ms to seconds per layer
@@ -920,6 +1036,7 @@ def _run_mla_benchmark_batched(
     configs_with_params: list[tuple],  # [(config, threshold, num_splits), ...]
     index_topk: int = 2048,
     prefill_backend: str | None = None,
+    sparse_mla_force_mqa: bool = False,
     output_scale: float | None = None,
     fuse_quant_op: bool = False,
 ) -> list[BenchmarkResult]:
@@ -940,6 +1057,8 @@ def _run_mla_benchmark_batched(
         index_topk: Topk value for sparse MLA backends (default 2048)
         prefill_backend: Prefill backend name (e.g., "fa3", "fa4").
             When set, forces the specified FlashAttention version for prefill.
+        sparse_mla_force_mqa: If True, forces all sparse MLA tokens through
+            forward_mqa (even prefill tokens).
 
     Returns:
         List of BenchmarkResult objects
@@ -980,21 +1099,41 @@ def _run_mla_benchmark_batched(
         sum(r.q_len for r in parse_batch_spec(cfg.batch_spec))
         for cfg, *_ in configs_with_params
     )
+    max_model_len = max(
+        max_total_q,
+        max(
+            getattr(cfg, "max_model_len", None) or 32768
+            for cfg, *_ in configs_with_params
+        ),
+    )
 
     # Create and set vLLM config for MLA (reused across all benchmarks)
     vllm_config = create_minimal_vllm_config(
         model_name="deepseek-v3",  # Used only for model path
         block_size=block_size,
         max_num_batched_tokens=max_total_q,
+        max_model_len=max_model_len,
         mla_dims=mla_dims,  # Use custom dims from config or default
         index_topk=index_topk if is_sparse else None,
         prefill_backend=prefill_backend,
         kv_cache_dtype=kv_cache_dtype,
+        sparse_mla_force_mqa=sparse_mla_force_mqa,
     )
 
     results = []
 
+    # Initialize workspace manager (needed by metadata builders)
+    from vllm.v1.worker.workspace import (
+        init_workspace_manager,
+        is_workspace_manager_initialized,
+    )
+
+    if not is_workspace_manager_initialized():
+        init_workspace_manager(device)
+
     with set_current_vllm_config(vllm_config):
+        _ensure_single_rank_model_parallel()
+
         # Create backend impl, layer, builder, and indexer (reused across benchmarks)
         impl, layer, builder_instance, indexer = _create_backend_impl(
             backend_cfg,
@@ -1040,9 +1179,20 @@ def _run_mla_benchmark_batched(
         for config, threshold, num_splits in configs_with_params:
             # Set threshold for this benchmark (FlashAttn/FlashMLA only)
             original_threshold = None
-            if threshold is not None and builder_instance:
+            effective_threshold = threshold
+            force_dense_mha = (
+                is_sparse
+                and getattr(config, "sparse_mla_mha_mode", "auto") == "dense"
+                and not getattr(config, "sparse_mla_force_mqa", False)
+            )
+            if force_dense_mha:
+                # Sparse MLA normally treats q_len <= 1 as decode. Use an
+                # impossible threshold so dense_mha benchmarks actually run
+                # the prefill/MHA path, including q_len=1 short extends.
+                effective_threshold = -1
+            if effective_threshold is not None and builder_instance:
                 original_threshold = builder_instance.reorder_batch_threshold
-                builder_instance.reorder_batch_threshold = threshold
+                builder_instance.reorder_batch_threshold = effective_threshold
 
             # Set num_splits for CUTLASS
             original_num_splits = None
@@ -1090,6 +1240,7 @@ def run_mla_benchmark(
     num_kv_splits: int | None = None,
     index_topk: int = 2048,
     prefill_backend: str | None = None,
+    sparse_mla_force_mqa: bool = False,
     output_scale: float | None = None,
     fuse_quant_op: bool = False,
 ) -> BenchmarkResult | list[BenchmarkResult]:
@@ -1111,6 +1262,8 @@ def run_mla_benchmark(
         index_topk: Topk value for sparse MLA backends (default 2048)
         prefill_backend: Prefill backend name (e.g., "fa3", "fa4").
             When set, forces the specified FlashAttention version for prefill.
+        sparse_mla_force_mqa: If True, forces all sparse MLA tokens through
+            forward_mqa (even prefill tokens).
         output_scale: Static per-tensor FP8 scale for prefill output (None = bf16).
         fuse_quant_op: With output_scale set, fuse the FP8 write into the prefill
             kernel vs a standalone post-quant kernel. See _run_single_benchmark.
@@ -1142,6 +1295,7 @@ def run_mla_benchmark(
         configs_with_params,
         index_topk,
         prefill_backend=prefill_backend,
+        sparse_mla_force_mqa=sparse_mla_force_mqa,
         output_scale=output_scale,
         fuse_quant_op=fuse_quant_op,
     )

benchmarks/kernels/bench_cp_gather_fp8.py

@@ -69,12 +69,11 @@ def make_inputs(total_tokens, num_reqs, block_size):
     # Output workspace
     dst = torch.zeros(total_tokens, HEAD_DIM, dtype=torch.bfloat16, device="cuda")
 
-    seq_lens_t = torch.tensor(seq_lens, dtype=torch.int32, device="cuda")
     workspace_starts_t = torch.tensor(
         workspace_starts, dtype=torch.int32, device="cuda"
     )
 
-    return cache, dst, block_table, seq_lens_t, workspace_starts_t
+    return cache, dst, block_table, workspace_starts_t
 
 
 def bench_scenario(label, num_reqs, total_tokens_list, save_path):
@@ -94,7 +93,7 @@ def bench_scenario(label, num_reqs, total_tokens_list, save_path):
         )
     )
     def bench_fn(total_tokens, provider, num_reqs):
-        cache, dst, block_table, seq_lens_t, ws_starts = make_inputs(
+        cache, dst, block_table, ws_starts = make_inputs(
             total_tokens, num_reqs, BLOCK_SIZE
         )
 
@@ -102,7 +101,7 @@ def bench_scenario(label, num_reqs, total_tokens_list, save_path):
 
         ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
             lambda: ops.cp_gather_and_upconvert_fp8_kv_cache(
-                cache, dst, block_table, seq_lens_t, ws_starts, num_reqs
+                cache, dst, block_table, ws_starts, num_reqs
             ),
             quantiles=quantiles,
             rep=500,

benchmarks/kernels/benchmark_k3_cutedsl_residual.py

@@ -0,0 +1,367 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Benchmark the Kimi-K3 latent MoE addmm against CuTe residual GEMM.
+
+The benchmark covers ``BF16[M, 3584] @ BF16[7168, 3584].T + BF16[M, 7168]``
+with FP32 accumulation and BF16 output. Both backends execute through CUDA
+Graph replay. Weights and residuals rotate across buffers exceeding L2 so the
+comparison models the full latent MoE projection-and-add path.
+"""
+
+from __future__ import annotations
+
+import argparse
+import dataclasses
+import importlib.util
+import json
+import math
+import statistics
+from collections.abc import Callable, Sequence
+from pathlib import Path
+from typing import Any
+
+import cutlass
+import cutlass.cute as cute
+import torch
+from cuda.bindings import driver as cuda
+from cuda.bindings.driver import CUstream
+from quack.compile_utils import make_fake_tensor
+
+N = 7168
+K = 3584
+
+
+@dataclasses.dataclass(frozen=True, slots=True)
+class Config:
+    block_size: int
+    outputs_per_block: int
+    k_unroll: int
+    vector_width: int = 8
+
+
+def parse_config(value: str) -> Config:
+    try:
+        parts = [int(part) for part in value.split(",")]
+    except ValueError as error:
+        raise argparse.ArgumentTypeError(
+            "config must be BLOCK,OUTPUTS,K_UNROLL[,VECTOR_WIDTH]"
+        ) from error
+    if len(parts) == 3:
+        return Config(*parts)
+    if len(parts) == 4:
+        return Config(*parts)
+    raise argparse.ArgumentTypeError(
+        "config must be BLOCK,OUTPUTS,K_UNROLL[,VECTOR_WIDTH]"
+    )
+
+
+def production_residual_config(m: int) -> Config | None:
+    """The measured Latent-MoE residual config for M, from the K3 table."""
+    from vllm.models.kimi_k3.nvidia.low_latency_gemm import KIMI_K3_PROJECTIONS
+
+    spec = KIMI_K3_PROJECTIONS.get((N, K))
+    config = spec.residual_config(m) if spec is not None else None
+    if config is None:
+        return None
+    return Config(
+        config.block_size,
+        config.outputs_per_block,
+        config.k_unroll,
+        config.vector_width,
+    )
+
+
+def candidate_configs(mode: str, selected: Config | None, m: int) -> list[Config]:
+    if mode == "selected":
+        if selected is not None:
+            return [selected]
+        # No explicit --config: fall back to the production table for this M.
+        config = production_residual_config(m)
+        return [config] if config is not None else []
+    if mode == "baseline":
+        return [Config(224, 4, 2)]
+    return [
+        Config(block_size, outputs_per_block, k_unroll, vector_width)
+        for vector_width in (4, 8)
+        for block_size in (32, 64, 128, 224, 448)
+        if block_size % 32 == 0 and K % (block_size * vector_width) == 0
+        for outputs_per_block in (1, 2, 4, 7, 8)
+        if N % outputs_per_block == 0
+        for k_unroll in (1, 2, 4)
+    ]
+
+
+def load_kernel_class(path: Path):
+    spec = importlib.util.spec_from_file_location("cute_skinny_device", path)
+    if spec is None or spec.loader is None:
+        raise RuntimeError(f"cannot load CuTe kernel from {path}")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module.CuteSkinnyGemm
+
+
+def stream() -> CUstream:
+    return CUstream(torch.cuda.current_stream().cuda_stream)
+
+
+def compile_kernel(kernel_class, m: int, config: Config, max_registers: int):
+    element_type = cutlass.BFloat16
+    n = cute.sym_int(divisibility=config.outputs_per_block)
+    k = cute.sym_int(divisibility=config.block_size * config.vector_width)
+    a = make_fake_tensor(element_type, (m, k), divisibility=config.vector_width)
+    b = make_fake_tensor(element_type, (n, k), divisibility=config.vector_width)
+    residual = make_fake_tensor(element_type, (m, n), divisibility=1)
+    c = make_fake_tensor(element_type, (m, n), divisibility=1)
+    kernel = kernel_class(
+        element_type=element_type,
+        num_rows=m,
+        block_size=config.block_size,
+        outputs_per_block=config.outputs_per_block,
+        vector_width=config.vector_width,
+        k_unroll=config.k_unroll,
+        has_residual=True,
+        use_pdl=True,
+    )
+    return cute.compile(
+        kernel,
+        a,
+        b,
+        residual,
+        c,
+        stream(),
+        options=(
+            "--enable-tvm-ffi --keep-cubin "
+            f"--ptxas-options -maxrregcount={max_registers} "
+            "--ptxas-options -lineinfo"
+        ),
+    )
+
+
+def resource_usage(compiled) -> dict[str, Any]:
+    executor = getattr(compiled, "_default_executor", None)
+    context = getattr(executor, "exec_context", None)
+    functions = getattr(context, "kernel_functions", None)
+    if not functions:
+        return {"resource_metrics_available": False}
+
+    def attribute(name, function) -> int:
+        error, value = cuda.cuFuncGetAttribute(name, function)
+        if error != cuda.CUresult.CUDA_SUCCESS:
+            raise RuntimeError(f"cuFuncGetAttribute failed with {error}")
+        return int(value)
+
+    registers = [
+        attribute(cuda.CUfunction_attribute.CU_FUNC_ATTRIBUTE_NUM_REGS, function)
+        for function in functions
+    ]
+    local_bytes = [
+        attribute(
+            cuda.CUfunction_attribute.CU_FUNC_ATTRIBUTE_LOCAL_SIZE_BYTES,
+            function,
+        )
+        for function in functions
+    ]
+    return {
+        "resource_metrics_available": True,
+        "registers_per_thread": max(registers, default=0),
+        "spill_bytes": max(local_bytes, default=0),
+    }
+
+
+def rotating_buffer_count(m: int, multiplier: float, limit: int) -> int:
+    properties = torch.cuda.get_device_properties(0)
+    bytes_per_pair = (N * K + m * N) * 2
+    target = math.ceil(multiplier * properties.L2_cache_size)
+    return max(2, min(limit, math.ceil(target / bytes_per_pair)))
+
+
+def graph_samples(
+    launch: Callable[[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor], None],
+    activation: torch.Tensor,
+    weights: Sequence[torch.Tensor],
+    residuals: Sequence[torch.Tensor],
+    repeats: int,
+    replays: int,
+) -> tuple[list[float], list[torch.Tensor]]:
+    outputs = [torch.empty_like(residual) for residual in residuals]
+    for weight, residual, output in zip(weights, residuals, outputs):
+        launch(activation, weight, residual, output)
+    torch.cuda.synchronize()
+
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        for weight, residual, output in zip(weights, residuals, outputs):
+            launch(activation, weight, residual, output)
+    for _ in range(20):
+        graph.replay()
+    torch.cuda.synchronize()
+
+    samples = []
+    for _ in range(repeats):
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+        start.record()
+        for _ in range(replays):
+            graph.replay()
+        end.record()
+        end.synchronize()
+        samples.append(start.elapsed_time(end) * 1000.0 / (replays * len(weights)))
+    return samples, outputs
+
+
+def summarize(samples: Sequence[float]) -> dict[str, Any]:
+    ordered = sorted(samples)
+
+    def percentile(fraction: float) -> float:
+        position = fraction * (len(ordered) - 1)
+        lower = math.floor(position)
+        upper = math.ceil(position)
+        if lower == upper:
+            return ordered[lower]
+        weight = position - lower
+        return ordered[lower] * (1.0 - weight) + ordered[upper] * weight
+
+    mean = statistics.mean(samples)
+    return {
+        "median_us": statistics.median(samples),
+        "p10_us": percentile(0.1),
+        "p90_us": percentile(0.9),
+        "mean_us": mean,
+        "cv_pct": statistics.pstdev(samples) / mean * 100.0,
+        "samples_us": list(samples),
+    }
+
+
+def correctness(
+    output: torch.Tensor,
+    activation: torch.Tensor,
+    weight: torch.Tensor,
+    residual: torch.Tensor,
+) -> dict[str, Any]:
+    actual = output.float()
+    reference = activation.float() @ weight.float().t() + residual.float()
+    error = (actual - reference).abs()
+    scaled_error = error / (reference.abs() + 1.0)
+    cosine = torch.nn.functional.cosine_similarity(
+        actual.flatten(), reference.flatten(), dim=0
+    ).item()
+    return {
+        "valid": cosine > 0.999,
+        "cosine": cosine,
+        "max_abs_error": error.max().item(),
+        "max_scaled_error": scaled_error.max().item(),
+    }
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--kernel", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument(
+        "--mode", choices=("baseline", "sweep", "selected"), default="baseline"
+    )
+    parser.add_argument("--config", type=parse_config)
+    parser.add_argument("--m", type=int, action="append")
+    parser.add_argument("--config-shard", type=int, default=0)
+    parser.add_argument("--num-config-shards", type=int, default=1)
+    parser.add_argument("--repeats", type=int, default=21)
+    parser.add_argument("--replays", type=int, default=200)
+    parser.add_argument("--cache-multiplier", type=float, default=3.0)
+    parser.add_argument("--max-buffers", type=int, default=32)
+    parser.add_argument("--max-registers", type=int, default=64)
+    args = parser.parse_args()
+
+    token_counts = args.m or list(range(1, 17))
+    if any(not 1 <= m <= 16 for m in token_counts):
+        raise ValueError("expected 1 <= M <= 16")
+    if not 0 <= args.config_shard < args.num_config_shards:
+        raise ValueError("config shard must be in [0, num_config_shards)")
+    torch.cuda.set_device(0)
+    if torch.cuda.get_device_capability() != (10, 3):
+        raise RuntimeError("this benchmark requires SM103")
+
+    kernel_class = load_kernel_class(args.kernel)
+    properties = torch.cuda.get_device_properties(0)
+    metadata = {
+        "device": properties.name,
+        "compute_capability": list(torch.cuda.get_device_capability()),
+        "torch_version": torch.__version__,
+        "cuda_version": torch.version.cuda,
+    }
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    with args.output.open("w", encoding="utf-8") as output_file:
+        for m in token_counts:
+            configs = candidate_configs(args.mode, args.config, m)
+            torch.manual_seed(20260722 + m)
+            count = rotating_buffer_count(m, args.cache_multiplier, args.max_buffers)
+            activation = torch.randn((m, K), device="cuda", dtype=torch.bfloat16)
+            weights = [
+                torch.randn((N, K), device="cuda", dtype=torch.bfloat16)
+                for _ in range(count)
+            ]
+            residuals = [
+                torch.randn((m, N), device="cuda", dtype=torch.bfloat16)
+                for _ in range(count)
+            ]
+            candidates: list[tuple[str, Config | None]] = [("cublas_addmm", None)]
+            candidates.extend(
+                ("cute_residual", config)
+                for index, config in enumerate(configs)
+                if index % args.num_config_shards == args.config_shard
+            )
+            for backend, config in candidates:
+                row: dict[str, Any] = {
+                    "m": m,
+                    "n": N,
+                    "k": K,
+                    "backend": backend,
+                    "mode": args.mode,
+                    "config": dataclasses.asdict(config) if config else {},
+                    "num_buffers": count,
+                    "cache_multiplier": args.cache_multiplier,
+                    **metadata,
+                }
+                try:
+                    if backend == "cublas_addmm":
+                        launch = lambda a, b, residual, c: torch.addmm(
+                            residual, a, b.t(), out=c
+                        )
+                    else:
+                        if config is None:
+                            raise AssertionError("missing CuTe config")
+                        compiled = compile_kernel(
+                            kernel_class, m, config, args.max_registers
+                        )
+                        launch = lambda a, b, residual, c, fn=compiled: fn(
+                            a, b, residual, c, stream()
+                        )
+                        row.update(resource_usage(compiled))
+                    samples, outputs = graph_samples(
+                        launch,
+                        activation,
+                        weights,
+                        residuals,
+                        args.repeats,
+                        args.replays,
+                    )
+                    row.update(
+                        correctness(outputs[0], activation, weights[0], residuals[0])
+                    )
+                    row.update(summarize(samples))
+                except Exception as error:  # noqa: BLE001
+                    row.update(
+                        {
+                            "valid": False,
+                            "error": f"{type(error).__name__}: {error}",
+                        }
+                    )
+                output_file.write(json.dumps(row, sort_keys=True) + "\n")
+                output_file.flush()
+                print(json.dumps(row, sort_keys=True), flush=True)
+
+            del activation, weights, residuals
+            torch.cuda.empty_cache()
+
+
+if __name__ == "__main__":
+    main()

benchmarks/kernels/benchmark_kimi_k3_latent_moe_tail.py

@@ -0,0 +1,806 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Benchmark the Kimi K3 latent-MoE tail and its up-projection kernels.
+
+The ``up-projection`` subcommand isolates the TP-local dynamic and static-M
+skinny GEMMs. It rotates weights through a working set larger than L2 to model
+successive model layers.
+
+The ``whole-tail`` subcommand measures the distributed operator. Its reference
+path includes two AllReduces, RMSNorm, the replicated up-projection, and the
+final add. CUDA-event samples report the slowest rank so cross-rank skew is
+included.
+
+Examples:
+
+.. code-block:: console
+
+    .venv/bin/python \
+      benchmarks/kernels/benchmark_kimi_k3_latent_moe_tail.py up-projection
+
+    torchrun --nproc-per-node=8 \
+      benchmarks/kernels/benchmark_kimi_k3_latent_moe_tail.py whole-tail
+
+For multi-node runs, launch one ``torchrun`` agent per node and use a shared
+rendezvous endpoint.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import os
+import statistics
+from collections.abc import Callable, Sequence
+from dataclasses import asdict
+from pathlib import Path
+from typing import Any
+
+import cutlass
+import cutlass.utils as utils
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+from cuda.bindings import driver as cuda
+
+from vllm.distributed import get_tp_group
+from vllm.distributed.parallel_state import (
+    init_distributed_environment,
+    initialize_model_parallel,
+    set_custom_all_reduce,
+)
+from vllm.model_executor.warmup.cutedsl_warmup import cutedsl_warmup
+from vllm.models.kimi_k3.nvidia.ops import latent_moe_tail
+from vllm.models.kimi_k3.nvidia.ops.cute_dsl.latent_moe_tail import (
+    fused_add_multicast_gemm,
+    fused_add_multicast_skinny_gemm,
+)
+
+HIDDEN_SIZE = 7168
+LATENT_SIZE = 3584
+RMS_EPS = 0.1
+MAX_NUM_TOKENS = 16
+MMA_TILER_MN = (64, 32)
+CLUSTER_SHAPE_MN = (1, 8)
+B_PRIME_STAGES = 2
+
+
+def parse_up_projection_config(
+    value: str,
+) -> fused_add_multicast_skinny_gemm.SkinnyConfig:
+    try:
+        values = [int(part) for part in value.split(",")]
+    except ValueError as error:
+        raise argparse.ArgumentTypeError(
+            "config must be BLOCK,OUTPUTS,K_UNROLL[,VECTOR_WIDTH[,PREFETCH_B]]"
+        ) from error
+    if len(values) in (3, 4):
+        return fused_add_multicast_skinny_gemm.SkinnyConfig(*values)
+    if len(values) == 5 and values[4] in (0, 1):
+        return fused_add_multicast_skinny_gemm.SkinnyConfig(
+            *values[:4],
+            prefetch_b_before_pdl=bool(values[4]),
+        )
+    raise argparse.ArgumentTypeError(
+        "config must be BLOCK,OUTPUTS,K_UNROLL"
+        "[,VECTOR_WIDTH[,PREFETCH_B]], where PREFETCH_B is 0 or 1"
+    )
+
+
+def parse_tail_skinny_config(
+    value: str,
+) -> tuple[int, fused_add_multicast_skinny_gemm.SkinnyConfig]:
+    try:
+        values = [int(part) for part in value.split(",")]
+    except ValueError as error:
+        raise argparse.ArgumentTypeError(
+            "config must be M,BLOCK,OUTPUTS,K_UNROLL[,VECTOR_WIDTH[,PREFETCH_B]]"
+        ) from error
+    if len(values) == 4:
+        num_tokens, *config = values
+        return num_tokens, fused_add_multicast_skinny_gemm.SkinnyConfig(*config)
+    if len(values) == 5:
+        num_tokens, *config = values
+        return num_tokens, fused_add_multicast_skinny_gemm.SkinnyConfig(*config)
+    if len(values) == 6 and values[5] in (0, 1):
+        num_tokens, block, outputs, unroll, vector_width, prefetch = values
+        return num_tokens, fused_add_multicast_skinny_gemm.SkinnyConfig(
+            block,
+            outputs,
+            unroll,
+            vector_width,
+            bool(prefetch),
+        )
+    raise argparse.ArgumentTypeError(
+        "config must be M,BLOCK,OUTPUTS,K_UNROLL"
+        "[,VECTOR_WIDTH[,PREFETCH_B]], where PREFETCH_B is 0 or 1"
+    )
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    subparsers = parser.add_subparsers(dest="scope", required=True)
+
+    up_projection = subparsers.add_parser(
+        "up-projection",
+        help="Benchmark the isolated TP-local up-projection kernels.",
+    )
+    up_projection.add_argument(
+        "--backend",
+        choices=("dynamic", "skinny", "both"),
+        default="both",
+    )
+    up_projection.add_argument("--tp-size", type=int, default=16)
+    up_projection.add_argument(
+        "--num-tokens",
+        type=int,
+        nargs="+",
+        default=[*range(1, 9), 16],
+    )
+    up_projection.add_argument(
+        "--skinny-config",
+        type=parse_up_projection_config,
+        action="append",
+        help="Benchmark a static-M config for every selected token count.",
+    )
+    up_projection.add_argument("--cache-multiplier", type=float, default=2.0)
+    up_projection.add_argument("--max-weights", type=int, default=64)
+    up_projection.add_argument("--warmup-replays", type=int, default=10)
+    up_projection.add_argument("--samples", type=int, default=31)
+    up_projection.add_argument("--output", type=Path)
+
+    whole_tail = subparsers.add_parser(
+        "whole-tail",
+        help="Benchmark the distributed latent-MoE tail operator.",
+    )
+    whole_tail.add_argument(
+        "--backend",
+        choices=("reference", "fused", "both"),
+        default="both",
+    )
+    whole_tail.add_argument(
+        "--num-tokens",
+        type=int,
+        nargs="+",
+        default=[1, 5, 8, 16],
+    )
+    whole_tail.add_argument("--warmup-replays", type=int, default=20)
+    whole_tail.add_argument("--samples", type=int, default=51)
+    whole_tail.add_argument(
+        "--skinny-max-num-tokens",
+        type=int,
+        nargs="+",
+        help="Override the fused operator's static-M cutoff; use 0 for dynamic-only.",
+    )
+    whole_tail.add_argument(
+        "--skinny-config",
+        type=parse_tail_skinny_config,
+        action="append",
+        help="Override one static-M config for tuning.",
+    )
+    whole_tail.add_argument("--output", type=Path)
+    return parser.parse_args()
+
+
+def percentile(samples: Sequence[float], fraction: float) -> float:
+    ordered = sorted(samples)
+    position = fraction * (len(ordered) - 1)
+    lower = math.floor(position)
+    upper = math.ceil(position)
+    if lower == upper:
+        return ordered[lower]
+    upper_weight = position - lower
+    return ordered[lower] * (1.0 - upper_weight) + ordered[upper] * upper_weight
+
+
+def summarize(samples_us: Sequence[float]) -> dict[str, Any]:
+    mean_us = statistics.mean(samples_us)
+    return {
+        "median_us": statistics.median(samples_us),
+        "p10_us": percentile(samples_us, 0.1),
+        "p90_us": percentile(samples_us, 0.9),
+        "mean_us": mean_us,
+        "cv_pct": statistics.pstdev(samples_us) / mean_us * 100.0,
+        "samples_us": list(samples_us),
+    }
+
+
+def rotating_weight_count(
+    shard_size: int,
+    cache_multiplier: float,
+    limit: int,
+) -> int:
+    properties = torch.cuda.get_device_properties(
+        torch.accelerator.current_device_index()
+    )
+    weight_bytes = shard_size * LATENT_SIZE * 2
+    target_bytes = math.ceil(properties.L2_cache_size * cache_multiplier)
+    return max(2, min(limit, math.ceil(target_bytes / weight_bytes)))
+
+
+def capture_up_projection_graph(
+    launches: Sequence[Callable[[], None]],
+) -> torch.cuda.CUDAGraph:
+    for launch in launches:
+        launch()
+    torch.accelerator.synchronize()
+
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        for launch in launches:
+            launch()
+    torch.accelerator.synchronize()
+    return graph
+
+
+def benchmark_up_projection_graph(
+    graph: torch.cuda.CUDAGraph,
+    *,
+    operations_per_replay: int,
+    warmup_replays: int,
+    samples: int,
+) -> dict[str, Any]:
+    for _ in range(warmup_replays):
+        graph.replay()
+    torch.accelerator.synchronize()
+
+    samples_us = []
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+    for _ in range(samples):
+        start.record()
+        graph.replay()
+        end.record()
+        end.synchronize()
+        samples_us.append(start.elapsed_time(end) * 1000.0 / operations_per_replay)
+    return summarize(samples_us)
+
+
+class DynamicKernel:
+    def __init__(
+        self,
+        shard_size: int,
+        mailbox: torch.Tensor,
+        shared_shard: torch.Tensor,
+    ) -> None:
+        self.shard_size = shard_size
+        self.mailbox = mailbox
+        self.mailbox_c = fused_add_multicast_gemm._as_cute(mailbox)
+        compile_latent = torch.empty(
+            (1, MAX_NUM_TOKENS, LATENT_SIZE),
+            dtype=torch.bfloat16,
+            device=mailbox.device,
+        )
+        compile_weight = torch.empty(
+            (1, shard_size, LATENT_SIZE),
+            dtype=torch.bfloat16,
+            device=mailbox.device,
+        )
+        cluster_size = math.prod(CLUSTER_SHAPE_MN)
+        max_active_clusters = utils.HardwareInfo().get_max_active_clusters(cluster_size)
+        self.compiled = fused_add_multicast_gemm.compile_kernel(
+            (MAX_NUM_TOKENS, shard_size, LATENT_SIZE, 1),
+            fused_add_multicast_gemm._as_cute(
+                compile_latent,
+                dynamic_m=True,
+            ),
+            fused_add_multicast_gemm._as_cute(compile_weight),
+            self.mailbox_c,
+            fused_add_multicast_gemm._as_cute(shared_shard),
+            HIDDEN_SIZE,
+            shard_size,
+            MMA_TILER_MN,
+            CLUSTER_SHAPE_MN,
+            max_active_clusters,
+            B_PRIME_STAGES,
+        )
+
+    def launch(
+        self,
+        latent: torch.Tensor,
+        weight: torch.Tensor,
+        shared_shard: torch.Tensor,
+    ) -> None:
+        stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
+        self.compiled(
+            fused_add_multicast_gemm._as_cute(
+                latent.unsqueeze(0),
+                dynamic_m=True,
+            ),
+            fused_add_multicast_gemm._as_cute(weight.unsqueeze(0)),
+            self.mailbox_c,
+            fused_add_multicast_gemm._as_cute(shared_shard),
+            cutlass.Int64(latent.shape[0]),
+            cutlass.Int64(self.mailbox.data_ptr()),
+            stream,
+        )
+
+
+class SkinnyKernel:
+    def __init__(
+        self,
+        num_tokens: int,
+        shard_size: int,
+        config: fused_add_multicast_skinny_gemm.SkinnyConfig,
+    ) -> None:
+        self.compiled = fused_add_multicast_skinny_gemm.compile_kernel(
+            num_rows=num_tokens,
+            latent_dim=LATENT_SIZE,
+            hidden_dim=HIDDEN_SIZE,
+            shard_dim=shard_size,
+            config=config,
+        )
+
+    def launch(
+        self,
+        latent: torch.Tensor,
+        weight: torch.Tensor,
+        shared_shard: torch.Tensor,
+        mailbox: torch.Tensor,
+    ) -> None:
+        self.compiled(
+            fused_add_multicast_skinny_gemm._as_cute(latent),
+            fused_add_multicast_skinny_gemm._as_cute(weight),
+            fused_add_multicast_skinny_gemm._as_cute(shared_shard),
+            cutlass.Int64(mailbox.data_ptr()),
+            cuda.CUstream(torch.cuda.current_stream().cuda_stream),
+        )
+
+
+def check_up_projection_output(
+    actual: torch.Tensor,
+    latent: torch.Tensor,
+    weight: torch.Tensor,
+    shared_shard: torch.Tensor,
+) -> None:
+    gemm = F.linear(latent.float(), weight.float()).to(torch.bfloat16)
+    expected = (gemm.float() + shared_shard.float()).to(torch.bfloat16)
+    torch.testing.assert_close(actual, expected, atol=8e-2, rtol=3e-2)
+
+
+def make_up_projection_launches(
+    launch: Callable[[torch.Tensor, torch.Tensor, torch.Tensor], None],
+    latent: torch.Tensor,
+    weights: Sequence[torch.Tensor],
+    shared_shard: torch.Tensor,
+) -> list[Callable[[], None]]:
+    return [
+        lambda weight=weight: launch(latent, weight, shared_shard) for weight in weights
+    ]
+
+
+def benchmark_up_projection(args: argparse.Namespace) -> None:
+    if args.tp_size <= 0 or HIDDEN_SIZE % args.tp_size:
+        raise ValueError("TP size must be positive and divide the hidden size")
+    if any(not 1 <= num_tokens <= MAX_NUM_TOKENS for num_tokens in args.num_tokens):
+        raise ValueError("--num-tokens values must be in [1, 16]")
+    if args.cache_multiplier <= 0 or args.max_weights <= 0:
+        raise ValueError("cache multiplier and max weights must be positive")
+    if args.warmup_replays < 0 or args.samples <= 0:
+        raise ValueError("warmup replays must be nonnegative and samples positive")
+
+    torch.accelerator.set_device_index(0)
+    device = torch.device("cuda", 0)
+    if torch.cuda.get_device_capability(device)[0] != 10:
+        raise RuntimeError("Kimi K3 latent-MoE tail requires SM100")
+
+    shard_size = HIDDEN_SIZE // args.tp_size
+    weight_count = rotating_weight_count(
+        shard_size,
+        args.cache_multiplier,
+        args.max_weights,
+    )
+    torch.manual_seed(20260726)
+    weights = [
+        torch.randn(
+            (shard_size, LATENT_SIZE),
+            dtype=torch.bfloat16,
+            device=device,
+        )
+        / LATENT_SIZE**0.5
+        for _ in range(weight_count)
+    ]
+    mailbox = torch.empty(
+        (1, MAX_NUM_TOKENS, HIDDEN_SIZE),
+        dtype=torch.bfloat16,
+        device=device,
+    )
+    shared = torch.randn(
+        (MAX_NUM_TOKENS, HIDDEN_SIZE),
+        dtype=torch.bfloat16,
+        device=device,
+    )
+    shared_shard = shared[:, :shard_size]
+    use_dynamic = args.backend in ("dynamic", "both")
+    use_skinny = args.backend in ("skinny", "both")
+    dynamic_kernel = (
+        DynamicKernel(shard_size, mailbox, shared_shard) if use_dynamic else None
+    )
+
+    results = []
+    for num_tokens in args.num_tokens:
+        latent = torch.randn(
+            (num_tokens, LATENT_SIZE),
+            dtype=torch.bfloat16,
+            device=device,
+        )
+        result: dict[str, Any] = {"num_tokens": num_tokens}
+        if dynamic_kernel is not None:
+            launches = make_up_projection_launches(
+                dynamic_kernel.launch,
+                latent,
+                weights,
+                shared_shard,
+            )
+            graph = capture_up_projection_graph(launches)
+            result["dynamic"] = benchmark_up_projection_graph(
+                graph,
+                operations_per_replay=len(launches),
+                warmup_replays=args.warmup_replays,
+                samples=args.samples,
+            )
+            check_up_projection_output(
+                mailbox[0, :num_tokens, :shard_size],
+                latent,
+                weights[-1],
+                shared_shard[:num_tokens],
+            )
+        if use_skinny:
+            configs = args.skinny_config or [
+                fused_add_multicast_skinny_gemm.config_for_m(
+                    num_tokens,
+                    shard_size,
+                )
+            ]
+            skinny_results = []
+            for config in configs:
+                skinny_kernel = SkinnyKernel(num_tokens, shard_size, config)
+
+                def launch_skinny(
+                    latent: torch.Tensor,
+                    weight: torch.Tensor,
+                    shared_shard: torch.Tensor,
+                    *,
+                    skinny_kernel: SkinnyKernel = skinny_kernel,
+                    num_tokens: int = num_tokens,
+                ) -> None:
+                    skinny_kernel.launch(
+                        latent,
+                        weight,
+                        shared_shard[:num_tokens],
+                        mailbox,
+                    )
+
+                launches = make_up_projection_launches(
+                    launch_skinny,
+                    latent,
+                    weights,
+                    shared_shard,
+                )
+                graph = capture_up_projection_graph(launches)
+                timing = benchmark_up_projection_graph(
+                    graph,
+                    operations_per_replay=len(launches),
+                    warmup_replays=args.warmup_replays,
+                    samples=args.samples,
+                )
+                check_up_projection_output(
+                    mailbox[0, :num_tokens, :shard_size],
+                    latent,
+                    weights[-1],
+                    shared_shard[:num_tokens],
+                )
+                skinny_results.append(
+                    {
+                        "config": asdict(config),
+                        **timing,
+                    }
+                )
+            result["skinny"] = skinny_results
+        results.append(result)
+
+    properties = torch.cuda.get_device_properties(device)
+    report = {
+        "scope": "up-projection",
+        "device": properties.name,
+        "compute_capability": list(torch.cuda.get_device_capability(device)),
+        "tp_size": args.tp_size,
+        "shard_size": shard_size,
+        "weight_count": weight_count,
+        "cache_multiplier": args.cache_multiplier,
+        "warmup_replays": args.warmup_replays,
+        "samples": args.samples,
+        "results": results,
+    }
+    rendered = json.dumps(report, indent=2)
+    print(rendered, flush=True)
+    if args.output is not None:
+        args.output.parent.mkdir(parents=True, exist_ok=True)
+        args.output.write_text(rendered + "\n", encoding="utf-8")
+
+
+def capture_tail_graph(
+    operation: Callable[[], torch.Tensor],
+    cpu_group: dist.ProcessGroup,
+) -> tuple[torch.cuda.CUDAGraph, torch.Tensor]:
+    for _ in range(3):
+        dist.barrier(group=cpu_group)
+        output = operation()
+    torch.accelerator.synchronize()
+
+    dist.barrier(group=cpu_group)
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        output = operation()
+    torch.accelerator.synchronize()
+    return graph, output
+
+
+def benchmark_tail_graph(
+    graph: torch.cuda.CUDAGraph,
+    *,
+    warmup_replays: int,
+    samples: int,
+    device_group: dist.ProcessGroup,
+    cpu_group: dist.ProcessGroup,
+) -> dict[str, Any]:
+    for _ in range(warmup_replays):
+        graph.replay()
+    torch.accelerator.synchronize()
+
+    dist.barrier(group=cpu_group)
+    starts = [torch.cuda.Event(enable_timing=True) for _ in range(samples + 1)]
+    ends = [torch.cuda.Event(enable_timing=True) for _ in range(samples + 1)]
+    for start, end in zip(starts, ends):
+        start.record()
+        graph.replay()
+        end.record()
+    torch.accelerator.synchronize()
+
+    samples_us = torch.tensor(
+        [start.elapsed_time(end) * 1000.0 for start, end in zip(starts, ends)],
+        dtype=torch.float64,
+        device=torch.accelerator.current_device_index(),
+    )
+    dist.all_reduce(samples_us, op=dist.ReduceOp.MAX, group=device_group)
+    return summarize(samples_us[1:].tolist())
+
+
+def make_inputs(
+    num_tokens: int,
+    rank: int,
+    device: torch.device,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    torch.manual_seed(20260726 + 100 * num_tokens + rank)
+    routed = torch.randn(
+        (num_tokens, LATENT_SIZE),
+        dtype=torch.bfloat16,
+        device=device,
+    ).mul_(0.01)
+    shared = torch.randn(
+        (num_tokens, HIDDEN_SIZE),
+        dtype=torch.bfloat16,
+        device=device,
+    )
+    return routed, shared
+
+
+def make_reference(
+    routed: torch.Tensor,
+    shared: torch.Tensor,
+    rms_weight: torch.Tensor,
+    up_weight: torch.Tensor,
+    device_group: dist.ProcessGroup,
+) -> Callable[[], torch.Tensor]:
+    routed_workspace = torch.empty_like(routed)
+    shared_workspace = torch.empty_like(shared)
+
+    def reference() -> torch.Tensor:
+        routed_workspace.copy_(routed)
+        dist.all_reduce(routed_workspace, group=device_group)
+        normalized = F.rms_norm(
+            routed_workspace,
+            (LATENT_SIZE,),
+            rms_weight,
+            RMS_EPS,
+        )
+        projected = F.linear(normalized, up_weight)
+        shared_workspace.copy_(shared)
+        dist.all_reduce(shared_workspace, group=device_group)
+        return projected.add(shared_workspace)
+
+    return reference
+
+
+def check_fused_output(
+    fused_output: torch.Tensor,
+    reference: Callable[[], torch.Tensor],
+    cpu_group: dist.ProcessGroup,
+) -> None:
+    dist.barrier(group=cpu_group)
+    expected = reference()
+    torch.testing.assert_close(fused_output, expected, atol=8e-2, rtol=3e-2)
+
+
+def benchmark_whole_tail(args: argparse.Namespace) -> None:
+    if any(not 1 <= num_tokens <= 16 for num_tokens in args.num_tokens):
+        raise ValueError("--num-tokens values must be in [1, 16]")
+    if args.warmup_replays < 0 or args.samples <= 0:
+        raise ValueError("warmup replays must be nonnegative and samples positive")
+    if args.skinny_max_num_tokens is not None and any(
+        not 0 <= cutoff <= 8 for cutoff in args.skinny_max_num_tokens
+    ):
+        raise ValueError("--skinny-max-num-tokens must be in [0, 8]")
+    skinny_configs = dict(args.skinny_config or ())
+    if len(skinny_configs) != len(args.skinny_config or ()):
+        raise ValueError("--skinny-config must not repeat an M value")
+    if any(not 1 <= num_tokens <= 8 for num_tokens in skinny_configs):
+        raise ValueError("--skinny-config M values must be in [1, 8]")
+    if not {"RANK", "WORLD_SIZE", "LOCAL_RANK"} <= os.environ.keys():
+        raise RuntimeError("launch this benchmark with torchrun")
+
+    rank = int(os.environ["RANK"])
+    world_size = int(os.environ["WORLD_SIZE"])
+    local_rank = int(os.environ["LOCAL_RANK"])
+    device = torch.device("cuda", local_rank)
+    torch.accelerator.set_device_index(device)
+    init_distributed_environment()
+    if world_size > 8:
+        set_custom_all_reduce(False)
+    initialize_model_parallel(tensor_model_parallel_size=world_size)
+    device_group = get_tp_group().device_group
+    cpu_group = dist.new_group(backend="gloo")
+
+    if torch.cuda.get_device_capability(device)[0] != 10:
+        raise RuntimeError("Kimi K3 latent-MoE tail requires SM100")
+
+    torch.manual_seed(20260726)
+    rms_weight = 1 + 0.1 * torch.randn(
+        LATENT_SIZE,
+        dtype=torch.bfloat16,
+        device=device,
+    )
+    up_weight = (
+        torch.randn(
+            (HIDDEN_SIZE, LATENT_SIZE),
+            dtype=torch.bfloat16,
+            device=device,
+        )
+        / LATENT_SIZE**0.5
+    )
+
+    use_reference = args.backend in ("reference", "both")
+    use_fused = args.backend in ("fused", "both")
+    fused_ops = []
+    if use_fused:
+        production_config_for_m = fused_add_multicast_skinny_gemm.config_for_m
+
+        def config_for_m(
+            num_rows: int,
+            shard_dim: int = 896,
+        ) -> fused_add_multicast_skinny_gemm.SkinnyConfig:
+            config = skinny_configs.get(num_rows)
+            if config is not None:
+                return config
+            return production_config_for_m(num_rows, shard_dim)
+
+        fused_add_multicast_skinny_gemm.config_for_m = config_for_m
+        cutoffs = args.skinny_max_num_tokens or [latent_moe_tail._SKINNY_MAX_NUM_TOKENS]
+        for cutoff in cutoffs:
+            latent_moe_tail._SKINNY_MAX_NUM_TOKENS = cutoff
+            latent_moe_tail.KimiK3LatentMoETailOp._instances.clear()
+            fused_ops.append(
+                (
+                    cutoff,
+                    latent_moe_tail.KimiK3LatentMoETailOp.initialize(
+                        hidden_size=HIDDEN_SIZE,
+                        latent_size=LATENT_SIZE,
+                        dtype=torch.bfloat16,
+                        device=device,
+                        rms_eps=RMS_EPS,
+                    ),
+                )
+            )
+        cutedsl_warmup()
+
+    results = []
+    for num_tokens in args.num_tokens:
+        routed, shared = make_inputs(num_tokens, rank, device)
+        reference = make_reference(
+            routed,
+            shared,
+            rms_weight,
+            up_weight,
+            device_group,
+        )
+        result: dict[str, Any] = {"num_tokens": num_tokens}
+        if use_reference:
+            reference_graph, _ = capture_tail_graph(reference, cpu_group)
+            result["reference"] = benchmark_tail_graph(
+                reference_graph,
+                warmup_replays=args.warmup_replays,
+                samples=args.samples,
+                device_group=device_group,
+                cpu_group=cpu_group,
+            )
+        for cutoff, fused_op in fused_ops:
+
+            def fused(
+                routed: torch.Tensor = routed,
+                shared: torch.Tensor = shared,
+                fused_op: latent_moe_tail.KimiK3LatentMoETailOp = fused_op,
+            ) -> torch.Tensor:
+                return fused_op(routed, shared, rms_weight, up_weight)
+
+            fused_graph, fused_output = capture_tail_graph(fused, cpu_group)
+            fused_key = "fused" if len(fused_ops) == 1 else f"fused_skinny_max_{cutoff}"
+            result[fused_key] = benchmark_tail_graph(
+                fused_graph,
+                warmup_replays=args.warmup_replays,
+                samples=args.samples,
+                device_group=device_group,
+                cpu_group=cpu_group,
+            )
+            check_fused_output(fused_output, reference, cpu_group)
+            if "reference" in result:
+                speedup = (
+                    result["reference"]["median_us"] / result[fused_key]["median_us"]
+                )
+                if len(fused_ops) == 1:
+                    result["speedup"] = speedup
+                else:
+                    result[f"{fused_key}_speedup"] = speedup
+        results.append(result)
+
+    properties = torch.cuda.get_device_properties(device)
+    report = {
+        "scope": "whole-tail",
+        "device": properties.name,
+        "compute_capability": list(torch.cuda.get_device_capability(device)),
+        "world_size": world_size,
+        "torch_version": torch.__version__,
+        "cuda_version": torch.version.cuda,
+        "warmup_replays": args.warmup_replays,
+        "samples": args.samples,
+        "skinny_max_num_tokens": [cutoff for cutoff, _ in fused_ops],
+        "skinny_configs": {
+            str(num_tokens): asdict(config)
+            for num_tokens, config in skinny_configs.items()
+        },
+        "timing_scope": {
+            "reference": (
+                "two input copies, two AllReduces, RMSNorm, full replicated "
+                "up-projection GEMM, and final add"
+            ),
+            "fused": (
+                "routed AllReduce/RMSNorm plus shared ReduceScatter, sharded "
+                "up-projection/multicast, and Lamport copy"
+            ),
+        },
+        "results": results,
+    }
+    if rank == 0:
+        rendered = json.dumps(report, indent=2)
+        print(rendered, flush=True)
+        if args.output is not None:
+            args.output.parent.mkdir(parents=True, exist_ok=True)
+            args.output.write_text(rendered + "\n", encoding="utf-8")
+
+    dist.barrier(group=cpu_group)
+
+
+def main() -> None:
+    args = parse_args()
+    if args.scope == "up-projection":
+        benchmark_up_projection(args)
+        return
+
+    from vllm.config import VllmConfig, set_current_vllm_config
+
+    with set_current_vllm_config(VllmConfig()):
+        benchmark_whole_tail(args)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/kernels/benchmark_kimi_k3_sp_collectives.py

@@ -0,0 +1,239 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+import json
+import os
+import statistics
+from collections.abc import Callable
+
+import torch
+import torch.distributed as dist
+
+import vllm._custom_ops as ops
+from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--tokens", type=int, nargs="+", default=[8, 32, 128, 1024])
+    parser.add_argument("--hidden-size", type=int, default=7168)
+    parser.add_argument("--graph-repeats", type=int, default=20)
+    parser.add_argument("--warmup-replays", type=int, default=5)
+    parser.add_argument("--samples", type=int, default=15)
+    return parser.parse_args()
+
+
+def capture_graph(op: Callable[[], None], repeats: int) -> torch.cuda.CUDAGraph:
+    stream = torch.cuda.Stream()
+    stream.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(stream):
+        for _ in range(3):
+            op()
+    stream.synchronize()
+
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph, stream=stream):
+        for _ in range(repeats):
+            op()
+    torch.cuda.current_stream().wait_stream(stream)
+    return graph
+
+
+def max_rank_graph_time(
+    graph: torch.cuda.CUDAGraph,
+    repeats: int,
+    warmup_replays: int,
+    samples: int,
+    device_group: dist.ProcessGroup,
+    cpu_group: dist.ProcessGroup,
+) -> float:
+    for _ in range(warmup_replays):
+        graph.replay()
+    torch.accelerator.synchronize()
+
+    timings = []
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+    for _ in range(samples):
+        dist.barrier(group=cpu_group)
+        start.record()
+        graph.replay()
+        end.record()
+        end.synchronize()
+        elapsed = torch.tensor(
+            start.elapsed_time(end) / repeats,
+            dtype=torch.float64,
+            device=torch.accelerator.current_device_index(),
+        )
+        dist.all_reduce(elapsed, op=dist.ReduceOp.MAX, group=device_group)
+        timings.append(elapsed.item())
+    return statistics.median(timings)
+
+
+def check_outputs(
+    comm: CustomAllreduce,
+    local: torch.Tensor,
+    reduce_input: torch.Tensor,
+    device_group: dist.ProcessGroup,
+) -> None:
+    expected_gather = torch.empty(
+        (local.shape[0] * dist.get_world_size(), local.shape[1]),
+        dtype=local.dtype,
+        device=local.device,
+    )
+    dist.all_gather_into_tensor(expected_gather, local, group=device_group)
+    gathered = comm.custom_all_gather(local)
+    assert gathered is not None
+    torch.testing.assert_close(gathered, expected_gather)
+
+    expected_scatter = torch.empty_like(local)
+    dist.reduce_scatter_tensor(
+        expected_scatter,
+        reduce_input.clone(),
+        group=device_group,
+    )
+    scattered = comm.custom_reduce_scatter(reduce_input)
+    assert scattered is not None
+    torch.testing.assert_close(scattered, expected_scatter)
+
+
+def benchmark_shape(
+    comm: CustomAllreduce,
+    global_tokens: int,
+    hidden_size: int,
+    graph_repeats: int,
+    warmup_replays: int,
+    samples: int,
+    device_group: dist.ProcessGroup,
+    cpu_group: dist.ProcessGroup,
+) -> dict[str, float | int]:
+    world_size = dist.get_world_size()
+    rank = dist.get_rank()
+    padded_tokens = (global_tokens + world_size - 1) // world_size * world_size
+    local_tokens = padded_tokens // world_size
+    local = torch.full(
+        (local_tokens, hidden_size),
+        rank + 1,
+        dtype=torch.bfloat16,
+        device=torch.accelerator.current_device_index(),
+    )
+    reduce_input = torch.full(
+        (padded_tokens, hidden_size),
+        rank + 1,
+        dtype=torch.bfloat16,
+        device=local.device,
+    )
+    check_outputs(comm, local, reduce_input, device_group)
+
+    custom_gather_out = torch.empty(
+        (padded_tokens, hidden_size),
+        dtype=local.dtype,
+        device=local.device,
+    )
+    custom_scatter_out = torch.empty_like(local)
+    nccl_gather_out = torch.empty_like(custom_gather_out)
+    nccl_scatter_out = torch.empty_like(local)
+
+    def custom_ag() -> None:
+        ops.mnnvl_lamport_all_gather(
+            comm._ptr,
+            local,
+            custom_gather_out,
+            comm.mnnvl_lamport_ag_local_ptr,
+            comm.mnnvl_lamport_ag_multicast_ptr,
+            comm.mnnvl_lamport_ag_epoch_ptr,
+            comm.mnnvl_buffer_size,
+        )
+
+    def custom_rs() -> None:
+        ops.mnnvl_lamport_reduce_scatter(
+            comm._ptr,
+            reduce_input,
+            custom_scatter_out,
+            comm.mnnvl_lamport_rs_local_ptr,
+            comm.mnnvl_lamport_rs_epoch_ptr,
+            comm.mnnvl_buffer_size,
+        )
+
+    def nccl_ag() -> None:
+        dist.all_gather_into_tensor(nccl_gather_out, local, group=device_group)
+
+    def nccl_rs() -> None:
+        dist.reduce_scatter_tensor(
+            nccl_scatter_out,
+            reduce_input,
+            group=device_group,
+        )
+
+    graphs = {
+        "custom_ag_us": capture_graph(custom_ag, graph_repeats),
+        "nccl_ag_us": capture_graph(nccl_ag, graph_repeats),
+        "custom_rs_us": capture_graph(custom_rs, graph_repeats),
+        "nccl_rs_us": capture_graph(nccl_rs, graph_repeats),
+    }
+    times = {
+        name: max_rank_graph_time(
+            graph,
+            graph_repeats,
+            warmup_replays,
+            samples,
+            device_group,
+            cpu_group,
+        )
+        * 1000
+        for name, graph in graphs.items()
+    }
+    torch.testing.assert_close(custom_gather_out, nccl_gather_out)
+    torch.testing.assert_close(custom_scatter_out, nccl_scatter_out)
+    return {
+        "global_tokens": global_tokens,
+        "padded_tokens": padded_tokens,
+        "local_bytes": local.nbytes,
+        "full_bytes": reduce_input.nbytes,
+        **times,
+        "ag_speedup": times["nccl_ag_us"] / times["custom_ag_us"],
+        "rs_speedup": times["nccl_rs_us"] / times["custom_rs_us"],
+    }
+
+
+def main() -> None:
+    args = parse_args()
+    local_rank = int(os.environ["LOCAL_RANK"])
+    torch.accelerator.set_device_index(local_rank)
+    dist.init_process_group("nccl")
+    device_group = dist.group.WORLD
+    cpu_group = dist.new_group(backend="gloo")
+
+    comm = CustomAllreduce(
+        group=cpu_group,
+        device=torch.device("cuda", local_rank),
+    )
+    assert not comm.disabled
+    assert comm.world_size == 16
+    assert comm.mnnvl_only
+    assert comm.mnnvl_multicast_ptr
+
+    results = [
+        benchmark_shape(
+            comm,
+            tokens,
+            args.hidden_size,
+            args.graph_repeats,
+            args.warmup_replays,
+            args.samples,
+            device_group,
+            cpu_group,
+        )
+        for tokens in args.tokens
+    ]
+    if dist.get_rank() == 0:
+        print(json.dumps(results, indent=2), flush=True)
+
+    comm.close()
+    dist.destroy_process_group(cpu_group)
+    dist.destroy_process_group()
+
+
+if __name__ == "__main__":
+    main()

benchmarks/kernels/benchmark_rdna_hybrid_w4a16_gemm.py

@@ -0,0 +1,201 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Benchmark the RDNAHybridW4A16LinearKernel across decode and prefill shapes.
+
+Usage:
+    python benchmark_int4_gemm.py
+    python benchmark_int4_gemm.py --models Qwen/Qwen3-4B
+    python benchmark_int4_gemm.py --group-size 128
+"""
+
+import argparse
+import copy
+import itertools
+import os
+
+import torch
+
+from vllm.triton_utils import triton
+
+# ---------------------------------------------------------------------------
+# Weight shapes: [K, N], TP_SPLIT_DIM
+# ---------------------------------------------------------------------------
+WEIGHT_SHAPES = {
+    "Qwen/Qwen3-4B": [
+        ([2560, 3840], 1),  # qkv_proj
+        ([2560, 2560], 0),  # o_proj
+        ([2560, 19456], 1),  # gate_up_proj
+        ([9728, 2560], 0),  # down_proj
+    ],
+    "Qwen/Qwen2.5-7B-Instruct": [
+        ([3584, 4608], 1),
+        ([3584, 3584], 0),
+        ([3584, 37888], 1),
+        ([18944, 3584], 0),
+    ],
+    "trymirai/SmolLM2-1.7B-Instruct-AWQ": [
+        ([2048, 6144], 1),  # qkv_proj
+        ([2048, 2048], 0),  # o_proj
+        ([2048, 16384], 1),  # gate_up_proj
+        ([8192, 2048], 0),  # down_proj
+    ],
+    "RedHatAI/Qwen3-8B-quantized.w4a16": [
+        ([4096, 6144], 1),  # qkv_proj
+        ([4096, 4096], 0),  # o_proj
+        ([4096, 24576], 1),  # gate_up_proj
+        ([12288, 4096], 0),  # down_proj
+    ],
+}
+
+
+# ---------------------------------------------------------------------------
+# Weight packing
+# ---------------------------------------------------------------------------
+def prepare_hybrid_weights(K, N, group_size, device="cuda"):
+    """Create random weights for benchmarking.
+
+    Returns (w_q_skinny, w_s_skinny, w_fp16, w_zp). The triton path derives
+    its int32 view from w_q_skinny, so no separate int32 buffer is returned.
+    """
+    num_groups = K // group_size
+
+    # Random packed weights — actual values don't matter for throughput
+    w_q_skinny_i32 = torch.randint(
+        0, 2**31, (N, K // 8), dtype=torch.int32, device=device
+    )
+    w_q_skinny = w_q_skinny_i32.view(torch.int8).contiguous()
+    w_s_skinny = torch.randn(N, num_groups, dtype=torch.float16, device=device) * 0.01
+
+    # Raw per-group zero-points for asymmetric benchmarks
+    w_zp = torch.randint(0, 16, (N, num_groups), dtype=torch.int32, device=device).to(
+        torch.float16
+    )
+
+    # FP16 baseline for F.linear
+    w_fp16 = torch.randn(N, K, dtype=torch.float16, device=device) * 0.01
+
+    return w_q_skinny, w_s_skinny, w_fp16, w_zp
+
+
+# ---------------------------------------------------------------------------
+# Benchmark
+# ---------------------------------------------------------------------------
+PROVIDERS = ["torch-fp16", "hybrid-w4a16", "hybrid-w4a16-zp"]
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096],
+        x_log=False,
+        line_arg="provider",
+        line_vals=PROVIDERS,
+        line_names=PROVIDERS,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="FP16 vs Hybrid W4A16",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K, group_size, weights):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.float16
+    a = torch.randn((M, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-fp16":
+        w_fp16 = weights["w_fp16"]
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, w_fp16),
+            quantiles=quantiles,
+        )
+    elif provider in ("hybrid-w4a16", "hybrid-w4a16-zp"):
+        from vllm.model_executor.kernels.linear.mixed_precision import (
+            rdna_hybrid_w4a16 as _k,
+        )
+
+        _rdna_hybrid_w4a16_apply_impl = _k._rdna_hybrid_w4a16_apply_impl
+        from vllm.utils.platform_utils import num_compute_units
+
+        w = weights
+        cu_count = num_compute_units()
+        use_zp = provider == "hybrid-w4a16-zp"
+
+        def run():
+            return _rdna_hybrid_w4a16_apply_impl(
+                a,
+                w["w_q_skinny"],
+                w["w_s_skinny"],
+                w["w_zp"] if use_zp else None,
+                None,  # bias
+                cu_count,
+                group_size,
+            )
+
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            run,
+            quantiles=quantiles,
+        )
+    else:
+        return 0.0, 0.0, 0.0
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    KN_model_names = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            KN_model_names.append(KN)
+    return KN_model_names
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Benchmark RDNAHybridW4A16LinearKernel"
+    )
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["Qwen/Qwen3-4B"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    parser.add_argument("--group-size", type=int, default=128)
+    parser.add_argument("--save-path", type=str, default=None)
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        group_size = args.group_size
+        print(f"\n{'=' * 70}")
+        print(f"{model}, N={N} K={K}, group_size={group_size}")
+        print(f"{'=' * 70}")
+
+        w_q_skinny, w_s_skinny, w_fp16, w_zp = prepare_hybrid_weights(K, N, group_size)
+
+        weights = {
+            "w_q_skinny": w_q_skinny,
+            "w_s_skinny": w_s_skinny,
+            "w_fp16": w_fp16,
+            "w_zp": w_zp,
+        }
+
+        save_path = args.save_path or f"bench_int4_res_n{N}_k{K}"
+        os.makedirs(save_path, exist_ok=True)
+        benchmark.run(
+            print_data=True,
+            show_plots=False,
+            save_path=save_path,
+            N=N,
+            K=K,
+            group_size=group_size,
+            weights=weights,
+        )
+
+    print("\nBenchmark finished!")

benchmarks/kernels/benchmark_relu_squared.py

@@ -0,0 +1,108 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Benchmark ReLUSquaredActivation: custom CUDA kernel vs forward_native, both
+# eager and under torch.compile (Inductor fuses relu+square into one kernel).
+
+import itertools
+
+import torch
+import torch.nn.functional as F
+
+import vllm.model_executor.layers.activation  # noqa: F401
+from vllm.benchmarks.lib.utils import default_vllm_config
+from vllm.triton_utils import triton
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed
+
+# Capped so the largest tensor stays under 2**31 elements: the shared activation
+# kernel computes the per-token pointer offset (blockIdx.x * d) in 32-bit, which
+# overflows for tensors with >2**32 elements. Realistic token counts are well
+# below this; the kernel-vs-native gap is already clear at these sizes.
+batch_size_range = [1, 16, 128]
+seq_len_range = [1, 16, 64, 1024]
+intermediate_size = [3072, 9728, 12288]
+configs = list(itertools.product(batch_size_range, seq_len_range, intermediate_size))
+
+
+@default_vllm_config()
+def benchmark_relu_squared(
+    batch_size: int,
+    seq_len: int,
+    intermediate_size: int,
+    provider: str,
+    dtype: torch.dtype,
+):
+    device = "cuda"
+    num_tokens = batch_size * seq_len
+    set_random_seed(42)
+    torch.set_default_device(device)
+
+    x = torch.randn(num_tokens, intermediate_size, dtype=dtype, device=device)
+    out = torch.empty_like(x)
+
+    def native(x: torch.Tensor) -> torch.Tensor:
+        return torch.square(F.relu(x))
+
+    # Verify the custom kernel matches the native implementation before timing.
+    ref = native(x)
+    torch.ops._C.relu_squared(out, x)
+    torch.testing.assert_close(out, ref)
+
+    if provider == "custom":
+        # Custom CUDA kernel — single fused kernel.
+        fn = lambda: torch.ops._C.relu_squared(out, x)
+    elif provider == "native":
+        # forward_native, eager — relu and square as separate ops.
+        fn = lambda: native(x)
+    elif provider == "native_compiled":
+        # forward_native under torch.compile — Inductor fuses relu+square.
+        # This is the real production baseline (custom ops are off when
+        # Inductor is enabled), so it is the comparison reviewers care about.
+        compiled = torch.compile(native)
+        compiled(x)  # warm up / trigger compilation before timing
+        fn = lambda: compiled(x)
+
+    ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+        fn, quantiles=[0.5, 0.2, 0.8]
+    )
+    return ms, max_ms, min_ms
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="Benchmark ReLUSquaredActivation: custom kernel vs native."
+    )
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        choices=["half", "bfloat16", "float"],
+        default="bfloat16",
+    )
+    args = parser.parse_args()
+
+    dtype = STR_DTYPE_TO_TORCH_DTYPE[args.dtype]
+
+    perf_report = triton.testing.perf_report(
+        triton.testing.Benchmark(
+            x_names=["batch_size", "seq_len", "intermediate_size"],
+            x_vals=configs,
+            line_arg="provider",
+            line_vals=["custom", "native_compiled", "native"],
+            line_names=[
+                "Custom Kernel",
+                "Native (torch.compile)",
+                "Native (eager)",
+            ],
+            styles=[("blue", "-"), ("green", "-"), ("red", "-")],
+            ylabel="ms",
+            plot_name="relu_squared-eager-performance",
+            args={},
+        )
+    )
+
+    perf_report(
+        lambda batch_size, seq_len, intermediate_size, provider: benchmark_relu_squared(
+            batch_size, seq_len, intermediate_size, provider, dtype
+        )
+    ).run(print_data=True)

benchmarks/replayssm/e2e_decode_speedup.py

@@ -0,0 +1,267 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""End-to-end autoregressive decode benchmark: ReplaySSM vs the standard SSM kernel.
+
+Loads a hybrid Mamba2 model, replicates one prompt across the batch, and times a
+long greedy decode (CUDA graphs on) once with the standard kernel and once with
+ReplaySSM, then reports the per-step / throughput speedup. The two modes run in
+separate subprocesses so each gets a clean CUDA context.
+
+The FlashInfer FP4-MoE autotuner is disabled by default (it is unstable under
+CUDA-graph capture on the pre-release Blackwell FP4 path); pass
+--no-disable-flashinfer-autotune for non-FP4 models.
+
+Examples:
+    python e2e_decode_speedup.py --model-id nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
+    python e2e_decode_speedup.py --dtype auto --buffer-len 16 \
+        --model-id nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4   # B300 NVFP4
+"""
+
+import argparse
+import json
+import os
+import subprocess
+import sys
+import time
+
+DEFAULT_PROMPT = "My cat wrote all this CUDA code for a new language model and"
+
+MODE_LABEL = {"standard": "standard", "replayssm": "ReplaySSM"}
+
+
+def parse_args():
+    p = argparse.ArgumentParser(
+        description="E2E decode speedup: ReplaySSM vs the standard SSM kernel."
+    )
+    p.add_argument("--model-id", default="nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16")
+    p.add_argument("--prompt", default=DEFAULT_PROMPT)
+    p.add_argument("--batch-size", type=int, default=256)
+    p.add_argument("--num-steps", type=int, default=1000)
+    p.add_argument("--warmup-steps", type=int, default=128)
+    p.add_argument("--repeats", type=int, default=1)
+    p.add_argument(
+        "--buffer-len", type=int, default=16, help="ReplaySSM input-buffer length."
+    )
+    p.add_argument(
+        "--dtype",
+        default="bfloat16",
+        choices=["bfloat16", "float16", "float32", "auto"],
+    )
+    p.add_argument("--gpu-memory-utilization", type=float, default=0.9)
+    p.add_argument("--max-model-len", type=int, default=None)
+    p.add_argument(
+        "--disable-flashinfer-autotune",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+        help="Disable the FlashInfer FP4-MoE autotuner (default: on). "
+        "It is unstable under CUDA-graph capture on the "
+        "pre-release Blackwell FP4 path; pass "
+        "--no-disable-flashinfer-autotune for non-FP4 models.",
+    )
+    p.add_argument(
+        "--mamba-ssm-cache-dtype",
+        default="auto",
+        choices=["auto", "float32", "float16", "bfloat16"],
+        help="SSM state dtype (both modes). 'auto' = config-driven; "
+        "'float32' = fp32 state, 'bfloat16' = s16 state.",
+    )
+    p.add_argument(
+        "--baseline-ssm-config",
+        default="",
+        help="Pin the STANDARD baseline's SSM launch config as "
+        "'bsm,nw' via override_ssm_config (forces the in-process "
+        "engine so the override reaches the kernel). Empty = off.",
+    )
+    p.add_argument(
+        "--worker",
+        choices=["standard", "replayssm"],
+        default=None,
+        help=argparse.SUPPRESS,
+    )
+    return p.parse_args()
+
+
+def resolve_max_model_len(args) -> int:
+    if args.max_model_len is not None:
+        return args.max_model_len
+    return args.num_steps + 256
+
+
+def run_worker(args):
+    # override_ssm_config is a module global; it only reaches the model if the
+    # engine runs in-process (default V1 spawns a separate EngineCore). Force it.
+    if args.worker == "standard" and args.baseline_ssm_config:
+        os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
+
+    import torch
+
+    from vllm import LLM, SamplingParams
+
+    mode = args.worker
+    max_model_len = resolve_max_model_len(args)
+
+    llm_kwargs = dict(
+        model=args.model_id,
+        tensor_parallel_size=1,
+        dtype=args.dtype,
+        max_model_len=max_model_len,
+        trust_remote_code=True,
+        enable_prefix_caching=False,
+        enable_chunked_prefill=False,
+        max_num_seqs=args.batch_size,
+        max_num_batched_tokens=max(max_model_len, args.batch_size * 64),
+        enforce_eager=False,
+        disable_log_stats=True,
+        gpu_memory_utilization=args.gpu_memory_utilization,
+        # SSM state dtype (applies to both standard and ReplaySSM).
+        mamba_ssm_cache_dtype=args.mamba_ssm_cache_dtype,
+    )
+    if args.disable_flashinfer_autotune:
+        # FP4-MoE autotuner is unstable under CUDA-graph capture on Blackwell;
+        # re-enable (--no-disable-flashinfer-autotune) only for non-FP4 models.
+        llm_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
+    if mode == "replayssm":
+        llm_kwargs.update(use_replayssm=True, replayssm_buffer_len=args.buffer_len)
+
+    _ssm_cm = None
+    if mode == "standard" and args.baseline_ssm_config:
+        from vllm.model_executor.layers.mamba.ops.mamba_ssm import override_ssm_config
+
+        _bsm, _nw = (int(x) for x in args.baseline_ssm_config.split(","))
+        _ssm_cm = override_ssm_config((_bsm, _nw))
+        _ssm_cm.__enter__()  # active through LLM() graph capture + decode
+        print(
+            f"[{mode}] override_ssm_config -> (BLOCK_SIZE_M={_bsm}, num_warps={_nw})",
+            flush=True,
+        )
+
+    llm = LLM(**llm_kwargs)
+    prompts = [args.prompt] * args.batch_size
+
+    def timed_generate(n_tokens):
+        sp = SamplingParams(
+            n=1,
+            temperature=0.0,
+            ignore_eos=True,
+            min_tokens=n_tokens,
+            max_tokens=n_tokens,
+        )
+        if torch.accelerator.is_available():
+            torch.accelerator.synchronize()
+        t0 = time.perf_counter()
+        outs = llm.generate(prompts, sp, use_tqdm=False)
+        if torch.accelerator.is_available():
+            torch.accelerator.synchronize()
+        elapsed = time.perf_counter() - t0
+        produced = min(len(o.outputs[0].token_ids) for o in outs)
+        assert produced == n_tokens, f"expected {n_tokens} tokens, got {produced}"
+        return elapsed
+
+    timed_generate(args.warmup_steps)
+
+    best = None
+    for _ in range(args.repeats):
+        elapsed = timed_generate(args.num_steps)
+        tok_s = args.batch_size * args.num_steps / elapsed
+        per_step_ms = elapsed / args.num_steps * 1e3
+        print(
+            f"[{mode}] {elapsed:.3f}s  {tok_s:,.0f} tok/s  {per_step_ms:.3f} ms/step",
+            flush=True,
+        )
+        if best is None or elapsed < best["elapsed_s"]:
+            best = {
+                "mode": mode,
+                "elapsed_s": elapsed,
+                "tok_s": tok_s,
+                "per_step_ms": per_step_ms,
+            }
+
+    print("RESULT_JSON " + json.dumps(best), flush=True)
+    if _ssm_cm is not None:
+        _ssm_cm.__exit__(None, None, None)
+
+
+def run_one_mode(args, mode) -> dict:
+    cmd = [
+        sys.executable,
+        __file__,
+        "--worker",
+        mode,
+        "--model-id",
+        args.model_id,
+        "--prompt",
+        args.prompt,
+        "--batch-size",
+        str(args.batch_size),
+        "--num-steps",
+        str(args.num_steps),
+        "--warmup-steps",
+        str(args.warmup_steps),
+        "--repeats",
+        str(args.repeats),
+        "--buffer-len",
+        str(args.buffer_len),
+        "--dtype",
+        args.dtype,
+        "--gpu-memory-utilization",
+        str(args.gpu_memory_utilization),
+        "--mamba-ssm-cache-dtype",
+        args.mamba_ssm_cache_dtype,
+        "--baseline-ssm-config",
+        args.baseline_ssm_config,
+    ]
+    cmd.append(
+        "--disable-flashinfer-autotune"
+        if args.disable_flashinfer_autotune
+        else "--no-disable-flashinfer-autotune"
+    )
+    if args.max_model_len is not None:
+        cmd += ["--max-model-len", str(args.max_model_len)]
+
+    result = None
+    proc = subprocess.Popen(
+        cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, bufsize=1
+    )
+    for line in proc.stdout:
+        sys.stdout.write(line)
+        sys.stdout.flush()
+        if line.startswith("RESULT_JSON "):
+            result = json.loads(line[len("RESULT_JSON ") :])
+    proc.wait()
+    if proc.returncode != 0:
+        raise RuntimeError(f"mode '{mode}' worker exited with {proc.returncode}")
+    if result is None:
+        raise RuntimeError(f"mode '{mode}' produced no RESULT_JSON line")
+    return result
+
+
+def main():
+    args = parse_args()
+    if args.worker is not None:
+        run_worker(args)
+        return
+
+    print(
+        f"model={args.model_id}  batch_size={args.batch_size}  "
+        f"steps={args.num_steps}  buffer_len={args.buffer_len}  dtype={args.dtype}"
+    )
+
+    std = run_one_mode(args, "standard")
+    fla = run_one_mode(args, "replayssm")
+    speedup = std["per_step_ms"] / fla["per_step_ms"]
+
+    print()
+    header = f"{'mode':<10}{'ms/step':>12}{'tok/s':>16}{'wall (s)':>12}"
+    print(header)
+    print("-" * len(header))
+    for r in (std, fla):
+        print(
+            f"{MODE_LABEL[r['mode']]:<10}{r['per_step_ms']:>12.3f}"
+            f"{r['tok_s']:>16,.0f}{r['elapsed_s']:>12.3f}"
+        )
+    print("-" * len(header))
+    print(f"speedup (standard / ReplaySSM, per step): {speedup:.3f}x")
+
+
+if __name__ == "__main__":
+    main()

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/deepgemm.cmake

@@ -28,9 +28,9 @@ if(DEEPGEMM_SRC_DIR)
   message(STATUS "DeepGEMM using local DEEPGEMM_SRC_DIR: ${deepgemm_SOURCE_DIR}")
 else()
   # Keep in sync with tools/install_deepgemm.sh
-  set(_DEEPGEMM_UPSTREAM_REPO "https://github.com/deepseek-ai/DeepGEMM.git")
+  set(_DEEPGEMM_UPSTREAM_REPO "git@github.com:Inferact/DeepGEMM.git")
   # NOTE: This is currently targeting nv-dev branch due to sm120 support
-  set(_DEEPGEMM_UPSTREAM_TAG "a6b593d2826719dcf4892609af7b84ee23aaf32a")
+  set(_DEEPGEMM_UPSTREAM_TAG "f5a76426fa084087169693fd0cd815223576d6e9")
 
   set(_deepgemm_fc_root "${FETCHCONTENT_BASE_DIR}")
   if(NOT _deepgemm_fc_root)
@@ -68,6 +68,9 @@ endif()
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.9)
     list(APPEND DEEPGEMM_SUPPORT_ARCHS "10.0f")
+    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.4)
+      list(APPEND DEEPGEMM_SUPPORT_ARCHS "10.7f")
+    endif()
   else()
     list(APPEND DEEPGEMM_SUPPORT_ARCHS "10.0a")
   endif()

cmake/external_projects/flashkda.cmake

@@ -0,0 +1,74 @@
+include(FetchContent)
+
+if(DEFINED ENV{FLASH_KDA_SRC_DIR})
+  set(FLASH_KDA_SRC_DIR $ENV{FLASH_KDA_SRC_DIR})
+endif()
+
+if(FLASH_KDA_SRC_DIR)
+  FetchContent_Declare(
+    flashkda
+    SOURCE_DIR ${FLASH_KDA_SRC_DIR}
+  )
+else()
+  FetchContent_Declare(
+    flashkda
+    GIT_REPOSITORY git@github.com:Inferact/FlashKDA.git
+    GIT_TAG a3e42bbbece3bb38f7c426b880315294a336e82f
+    GIT_PROGRESS TRUE
+    GIT_SUBMODULES cutlass
+  )
+endif()
+
+FetchContent_MakeAvailable(flashkda)
+message(STATUS "FlashKDA is available at ${flashkda_SOURCE_DIR}")
+
+set(FLASH_KDA_SUPPORT_ARCHS)
+if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0)
+  list(APPEND FLASH_KDA_SUPPORT_ARCHS "9.0a")
+endif()
+if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+  list(APPEND FLASH_KDA_SUPPORT_ARCHS "10.0f" "12.0f")
+elseif(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.9)
+  list(APPEND FLASH_KDA_SUPPORT_ARCHS "10.0a" "10.3a" "12.0a")
+endif()
+
+cuda_archs_loose_intersection(
+  FLASH_KDA_ARCHS "${FLASH_KDA_SUPPORT_ARCHS}" "${CUDA_ARCHS}")
+
+if(FLASH_KDA_ARCHS)
+  message(STATUS "FlashKDA CUDA architectures: ${FLASH_KDA_ARCHS}")
+
+  set(FLASH_KDA_SOURCES
+    csrc/flashkda_registration.cpp
+    ${flashkda_SOURCE_DIR}/csrc/flash_kda.cpp
+    ${flashkda_SOURCE_DIR}/csrc/smxx/fwd_launch.cu)
+  set(FLASH_KDA_INCLUDES
+    ${flashkda_SOURCE_DIR}/csrc
+    ${flashkda_SOURCE_DIR}/cutlass/include
+    ${flashkda_SOURCE_DIR}/cutlass/examples/common
+    ${flashkda_SOURCE_DIR}/cutlass/tools/util/include)
+
+  set_gencode_flags_for_srcs(
+    SRCS "${FLASH_KDA_SOURCES}"
+    CUDA_ARCHS "${FLASH_KDA_ARCHS}")
+
+  define_extension_target(
+    _flashkda_C
+    DESTINATION vllm
+    LANGUAGE ${VLLM_GPU_LANG}
+    SOURCES ${FLASH_KDA_SOURCES}
+    COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+    ARCHITECTURES ${VLLM_GPU_ARCHES}
+    INCLUDE_DIRECTORIES ${FLASH_KDA_INCLUDES}
+    USE_SABI 3
+    WITH_SOABI)
+
+  target_compile_options(_flashkda_C PRIVATE
+    $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API --expt-relaxed-constexpr --expt-extended-lambda --use_fast_math -O3>
+    $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
+else()
+  message(STATUS
+    "FlashKDA will not compile: CUDA >=12.0 and a supported architecture "
+    "(SM90, SM10x, or SM12x) are required")
+  add_custom_target(_flashkda_C)
+endif()

cmake/external_projects/flashmla.cmake

@@ -19,7 +19,7 @@ else()
   FetchContent_Declare(
         flashmla
         GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
-        GIT_TAG a6ec2ba7bd0a7dff98b3f4d3e6b52b159c48d78b
+        GIT_TAG a8f794d1251cbfd88a5011445dd5582289c727e4
         GIT_PROGRESS TRUE
         CONFIGURE_COMMAND ""
         BUILD_COMMAND ""
@@ -35,7 +35,7 @@ set(FLASHMLA_VENDOR_DIR "${CMAKE_SOURCE_DIR}/vllm/third_party/flashmla")
 file(MAKE_DIRECTORY "${FLASHMLA_VENDOR_DIR}")
 file(READ "${flashmla_SOURCE_DIR}/flash_mla/flash_mla_interface.py"
      FLASHMLA_INTERFACE_CONTENT)
-string(REPLACE "import flash_mla.cuda as flash_mla_cuda"
+string(REPLACE "flash_mla_cuda = torch.ops._flashmla_C"
                "import vllm._flashmla_C\nflash_mla_cuda = torch.ops._flashmla_C"
                FLASHMLA_INTERFACE_CONTENT
                "${FLASHMLA_INTERFACE_CONTENT}")
@@ -60,6 +60,9 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.9)
     # CUDA 12.9 has introduced "Family-Specific Architecture Features"
     # this supports all compute_10x family
     list(APPEND SUPPORT_ARCHS "10.0f")
+    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.4)
+        list(APPEND SUPPORT_ARCHS "10.7f")
+    endif()
 elseif(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
     list(APPEND SUPPORT_ARCHS "10.0a")
 endif()
@@ -72,7 +75,7 @@ if(FLASH_MLA_ARCHS)
     list(APPEND VLLM_FLASHMLA_GPU_FLAGS "--expt-relaxed-constexpr" "--expt-extended-lambda" "--use_fast_math")
 
     set(FlashMLA_SOURCES
-        ${flashmla_SOURCE_DIR}/csrc/torch_api.cpp
+        ${flashmla_SOURCE_DIR}/csrc/api/api.cpp
 
         # Misc kernels for decoding
         ${flashmla_SOURCE_DIR}/csrc/smxx/decode/get_decoding_sched_meta/get_decoding_sched_meta.cu
@@ -128,6 +131,7 @@ if(FLASH_MLA_ARCHS)
 
     set(FlashMLA_Extension_INCLUDES
         ${flashmla_SOURCE_DIR}/csrc
+        ${flashmla_SOURCE_DIR}/csrc/kerutils/include
         ${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/
         ${flashmla_SOURCE_DIR}/csrc/cutlass/include
         ${flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
@@ -152,15 +156,18 @@ if(FLASH_MLA_ARCHS)
         USE_SABI 3
         WITH_SOABI)
 
-    # Keep Stable ABI for the module, but *not* for CUDA/C++ files.
-    # This prevents Py_LIMITED_API from affecting nvcc and C++ compiles.
-    # Also enable C++20 for the FlashMLA sources (required for std::span, requires, etc.)
+    # Enable C++20 for the FlashMLA sources (required for std::span, requires, etc.)
     target_compile_options(_flashmla_C PRIVATE
-        $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
-        $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>
         $<$<COMPILE_LANGUAGE:CXX>:-std=c++20>
         $<$<COMPILE_LANGUAGE:CUDA>:-std=c++20>)
 
+    # _flashmla_C is now ABI-stable torch 2.11+
+    target_compile_definitions(_flashmla_C PRIVATE
+        TORCH_TARGET_VERSION=0x020B000000000000ULL)
+    if(VLLM_GPU_LANG STREQUAL "CUDA")
+        target_compile_definitions(_flashmla_C PRIVATE USE_CUDA)
+    endif()
+
     define_extension_target(
         _flashmla_extension_C
         DESTINATION vllm
@@ -172,15 +179,15 @@ if(FLASH_MLA_ARCHS)
         USE_SABI 3
         WITH_SOABI)
 
-    # Keep Stable ABI for the module, but *not* for CUDA/C++ files.
-    # This prevents Py_LIMITED_API from affecting nvcc and C++ compiles.
-    target_compile_options(_flashmla_extension_C PRIVATE
-        $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
-        $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
+    # _flashmla_extension_C is now ABI-stable w/ torch 2.11+
+    target_compile_definitions(_flashmla_extension_C PRIVATE
+        TORCH_TARGET_VERSION=0x020B000000000000ULL)
+    if(VLLM_GPU_LANG STREQUAL "CUDA")
+        target_compile_definitions(_flashmla_extension_C PRIVATE USE_CUDA)
+    endif()
 else()
     message(STATUS "FlashMLA will not compile: unsupported CUDA architecture ${CUDA_ARCHS}")
     # Create empty targets for setup.py on unsupported systems
     add_custom_target(_flashmla_C)
     add_custom_target(_flashmla_extension_C)
 endif()
-

cmake/external_projects/fmha_sm100.cmake

@@ -17,7 +17,7 @@ else()
   FetchContent_Declare(
     fmha_sm100
     GIT_REPOSITORY https://github.com/vllm-project/MSA.git
-    GIT_TAG 2e63ec37a0fc29bc20f39cd1a52e0f5affc33a73
+    GIT_TAG 890aaa1a37a598ad17ccff0827fea21540d381fa
     GIT_PROGRESS TRUE
     CONFIGURE_COMMAND ""
     BUILD_COMMAND ""

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)
@@ -55,7 +55,11 @@ message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}")
 
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
   cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0f" "${CUDA_ARCHS}")
-  cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.4)
+    cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f;10.7f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f" "${CUDA_ARCHS}")
+  endif()
 else()
   cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0a;12.1a" "${CUDA_ARCHS}")
   cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0a;10.3a" "${CUDA_ARCHS}")
@@ -125,8 +129,6 @@ 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 +141,11 @@ 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/tml_fa4.cmake

@@ -0,0 +1,50 @@
+include(FetchContent)
+
+if(DEFINED ENV{TML_FA4_SRC_DIR})
+  set(TML_FA4_SRC_DIR $ENV{TML_FA4_SRC_DIR})
+endif()
+
+if(TML_FA4_SRC_DIR)
+  FetchContent_Declare(
+    tml_fa4
+    SOURCE_DIR ${TML_FA4_SRC_DIR}
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND "")
+else()
+  FetchContent_Declare(
+    tml_fa4
+    GIT_REPOSITORY https://github.com/vllm-project/tml-fa4.git
+    GIT_TAG b206834606ed5b5f21f8eed6b0683f528ea9cf7d
+    GIT_PROGRESS TRUE
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND "")
+endif()
+
+FetchContent_GetProperties(tml_fa4)
+if(NOT tml_fa4_POPULATED)
+  FetchContent_Populate(tml_fa4)
+endif()
+message(STATUS "tml-fa4 is available at ${tml_fa4_SOURCE_DIR}")
+
+add_custom_target(tml_fa4)
+
+# Install into a private namespace so this implementation cannot shadow the
+# flash_attn package used by vLLM's standard attention backends.
+install(CODE "
+  file(GLOB_RECURSE TML_FA4_PY_FILES
+    \"${tml_fa4_SOURCE_DIR}/flash_attn/cute/*.py\")
+  foreach(SRC_FILE \${TML_FA4_PY_FILES})
+    file(RELATIVE_PATH REL_PATH
+      \"${tml_fa4_SOURCE_DIR}/flash_attn/cute\" \${SRC_FILE})
+    set(DST_FILE
+      \"\${CMAKE_INSTALL_PREFIX}/vllm/third_party/tml_fa4/\${REL_PATH}\")
+    get_filename_component(DST_DIR \${DST_FILE} DIRECTORY)
+    file(MAKE_DIRECTORY \${DST_DIR})
+    file(READ \${SRC_FILE} FILE_CONTENTS)
+    string(REPLACE
+      \"flash_attn.cute\"
+      \"vllm.third_party.tml_fa4\"
+      FILE_CONTENTS \"\${FILE_CONTENTS}\")
+    file(WRITE \${DST_FILE} \"\${FILE_CONTENTS}\")
+  endforeach()
+" COMPONENT tml_fa4)

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 b3964b1d8b95d8e8447435668ab169a2700bab65
+          GIT_TAG ed4b7342bc8f0489dd9b649d5288867e35fc6a32
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

@@ -396,14 +396,24 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   # match — e.g. SRC="12.0f" matches TGT="12.1a" since SM121 is in the SM12x
   # family. The output uses TGT's value to preserve the user's compilation flags.
   set(_CUDA_ARCHS)
+  # Resolve exact base matches before family fallbacks so a generic entry such
+  # as 10.0f cannot consume a 10.7 target that has a 10.7f source entry.
+  foreach(_arch ${_SRC_CUDA_ARCHS})
+    if(_arch MATCHES "[af]$")
+      string(REGEX REPLACE "[af]$" "" _base "${_arch}")
+      if("${_base}" IN_LIST _TGT_CUDA_ARCHS)
+        list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
+        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
+        list(APPEND _CUDA_ARCHS "${_arch}")
+      endif()
+    endif()
+  endforeach()
+
   foreach(_arch ${_SRC_CUDA_ARCHS})
     if(_arch MATCHES "[af]$")
       list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
       string(REGEX REPLACE "[af]$" "" _base "${_arch}")
-      if ("${_base}" IN_LIST TGT_CUDA_ARCHS)
-        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
-        list(APPEND _CUDA_ARCHS "${_arch}")
-      elseif("${_base}a" IN_LIST _TGT_CUDA_ARCHS)
+      if("${_base}a" IN_LIST _TGT_CUDA_ARCHS)
         list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}a")
         list(APPEND _CUDA_ARCHS "${_base}a")
       elseif("${_base}f" IN_LIST _TGT_CUDA_ARCHS)
@@ -487,7 +497,7 @@ endfunction()
 
 function(cuda_archs_sm90plus OUT_CUDA_ARCHS TGT_CUDA_ARCHS)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(_archs "9.0a;10.0f;11.0f;12.0f" "${TGT_CUDA_ARCHS}")
+    cuda_archs_loose_intersection(_archs "9.0a;10.0f;10.7f;11.0f;12.0f" "${TGT_CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(_archs "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${TGT_CUDA_ARCHS}")
   endif()

csrc/cache.h

@@ -67,9 +67,8 @@ void cp_gather_and_upconvert_fp8_kv_cache(
     torch::Tensor const& src_cache,         // [NUM_BLOCKS, BLOCK_SIZE, 656]
     torch::Tensor const& dst,               // [TOT_TOKENS, 576]
     torch::Tensor const& block_table,       // [BATCH, BLOCK_INDICES]
-    torch::Tensor const& seq_lens,          // [BATCH]
     torch::Tensor const& workspace_starts,  // [BATCH]
-    int64_t batch_size);
+    int64_t batch_size, std::optional<torch::Tensor> seq_starts = std::nullopt);
 
 // Indexer K quantization and cache function
 void indexer_k_quant_and_cache(

csrc/cpu/cpu_attn_vec.hpp

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

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/sgl-kernels/fla.cpp

@@ -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]

csrc/cpu/torch_bindings.cpp

@@ -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);
@@ -205,6 +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 {
@@ -508,6 +542,15 @@ 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)");
@@ -553,7 +596,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 #endif
 
   // fused moe
-#if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT))
+#if defined(__AVX512F__) || \
+    (defined(__aarch64__) && !defined(__APPLE__) && defined(ARM_BF16_SUPPORT))
   ops.def(
       "prepack_moe_weight(Tensor weight, Tensor(a1!) packed_weight, str isa) "
       "-> ()");
@@ -564,7 +608,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "bool skip_weighted, "
       "str act, str isa) -> ()");
   ops.impl("cpu_fused_moe", torch::kCPU, &cpu_fused_moe);
-#endif  // #if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT))
+#endif
   ops.def(
       "mla_decode_kvcache("
       "   Tensor! out, Tensor query, Tensor kv_cache,"
@@ -577,6 +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/custom_all_gather_reduce_scatter.cuh

@@ -0,0 +1,326 @@
+#pragma once
+
+#include "custom_collective_common.cuh"
+
+namespace vllm {
+
+constexpr int kMnnvlLamportAgThreads = 128;
+constexpr int kMnnvlLamportRsThreads = 256;
+constexpr int kMnnvlLamportConcurrentPollMaxPacks = 8192;
+
+using CopyPack = array_t<uint64_t, 2>;
+
+template <int ngpus>
+__global__ void __launch_bounds__(512, 1)
+    cross_device_all_gather(RankData* _dp, RankSignals sg, Signal* self_sg,
+                            CopyPack* __restrict__ result, int rank,
+                            int size_per_rank) {
+  auto dp = *_dp;
+  int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  int stride = gridDim.x * blockDim.x;
+  barrier_at_start<ngpus>(sg, self_sg, rank);
+#pragma unroll
+  for (int src_rank = 0; src_rank < ngpus; ++src_rank) {
+    auto src = reinterpret_cast<const CopyPack*>(dp.ptrs[src_rank]);
+    auto dst = result + src_rank * size_per_rank;
+    for (int idx = tid; idx < size_per_rank; idx += stride) {
+      dst[idx] = src[idx];
+    }
+  }
+  barrier_at_end<ngpus, true>(sg, self_sg, rank);
+}
+
+template <typename T, int ngpus>
+__global__ void __launch_bounds__(512, 1)
+    cross_device_reduce_scatter(RankData* _dp, RankSignals sg, Signal* self_sg,
+                                T* __restrict__ result, int rank,
+                                int size_per_rank) {
+  using P = typename packed_t<T>::P;
+  using A = typename packed_t<T>::A;
+  auto dp = *_dp;
+  auto offset = rank * size_per_rank;
+  barrier_at_start<ngpus>(sg, self_sg, rank);
+  for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size_per_rank;
+       idx += gridDim.x * blockDim.x) {
+    reinterpret_cast<P*>(result)[idx] =
+        packed_reduce<P, ngpus, A>((const P**)&dp.ptrs[0], offset + idx);
+  }
+  barrier_at_end<ngpus, true>(sg, self_sg, rank);
+}
+
+template <typename P>
+union LamportPack {
+  P packed;
+  uint32_t words[sizeof(P) / sizeof(uint32_t)];
+};
+
+template <typename P>
+DINLINE LamportPack<P> load_lamport_pack(const P* ptr) {
+  static_assert(sizeof(P) == 16);
+  LamportPack<P> value;
+#if !defined(USE_ROCM)
+  asm volatile("ld.volatile.global.v4.u32 {%0, %1, %2, %3}, [%4];"
+               : "=r"(value.words[0]), "=r"(value.words[1]),
+                 "=r"(value.words[2]), "=r"(value.words[3])
+               : "l"(ptr)
+               : "memory");
+#else
+  const volatile uint32_t* src = reinterpret_cast<const volatile uint32_t*>(ptr);
+#pragma unroll
+  for (int i = 0; i < sizeof(P) / sizeof(uint32_t); ++i) {
+    value.words[i] = src[i];
+  }
+#endif
+  return value;
+}
+
+template <typename P>
+DINLINE bool is_lamport_dirty(const LamportPack<P>& value) {
+#pragma unroll
+  for (int i = 0; i < sizeof(P) / sizeof(uint32_t); ++i) {
+    if (value.words[i] == 0x80000000U) return true;
+  }
+  return false;
+}
+
+template <typename P>
+DINLINE P lamport_sentinel() {
+  LamportPack<P> value;
+#pragma unroll
+  for (int i = 0; i < sizeof(P) / sizeof(uint32_t); ++i) {
+    value.words[i] = 0x80000000U;
+  }
+  return value.packed;
+}
+
+template <typename P>
+DINLINE P sanitize_lamport_payload(P packed) {
+  LamportPack<P> value{.packed = packed};
+#pragma unroll
+  for (int i = 0; i < sizeof(P) / sizeof(uint32_t); ++i) {
+    if (value.words[i] == 0x80000000U) value.words[i] = 0;
+  }
+  return value.packed;
+}
+
+template <typename P>
+DINLINE P wait_lamport_payload(const P* ptr) {
+  auto value = load_lamport_pack(ptr);
+  while (is_lamport_dirty(value)) value = load_lamport_pack(ptr);
+  return value.packed;
+}
+
+template <typename P, int ngpus>
+DINLINE void wait_lamport_payloads(const P* base, int rank, int rank_stride,
+                                   P local_value, P (&values)[ngpus]) {
+  bool ready[ngpus];
+#pragma unroll
+  for (int src_rank = 0; src_rank < ngpus; ++src_rank) {
+    ready[src_rank] = src_rank == rank;
+    if (src_rank == rank) values[src_rank] = local_value;
+  }
+
+  int remaining = ngpus - 1;
+  while (remaining != 0) {
+#pragma unroll
+    for (int src_rank = 0; src_rank < ngpus; ++src_rank) {
+      if (!ready[src_rank]) {
+        auto value = load_lamport_pack(base + src_rank * rank_stride);
+        if (!is_lamport_dirty(value)) {
+          values[src_rank] = value.packed;
+          ready[src_rank] = true;
+          --remaining;
+        }
+      }
+    }
+  }
+}
+
+template <typename P, typename A, int ngpus>
+DINLINE P reduce_lamport_payloads(const P* current_local, const P* packed_input,
+                                  int rank, int size_per_rank, int idx) {
+  P source_zero =
+      rank == 0 ? packed_input[idx] : wait_lamport_payload(current_local + idx);
+  A tmp = upcast(source_zero);
+#pragma unroll
+  for (int src_rank = 1; src_rank < ngpus; ++src_rank) {
+    P value = src_rank == rank
+                  ? packed_input[rank * size_per_rank + idx]
+                  : wait_lamport_payload(current_local +
+                                         src_rank * size_per_rank + idx);
+    packed_assign_add(tmp, upcast(value));
+  }
+  return sanitize_lamport_payload(downcast<P>(tmp));
+}
+
+DINLINE void lamport_cta_arrive(uint32_t* counter) {
+#if !defined(USE_ROCM)
+  if (threadIdx.x < 32) {
+    asm volatile("barrier.cta.sync 1, %0;" : : "r"(blockDim.x) : "memory");
+    if (threadIdx.x == 0) {
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000
+      asm volatile("red.async.release.global.gpu.add.u32 [%0], 1;"
+                   :
+                   : "l"(counter)
+                   : "memory");
+  #elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+      asm volatile("red.release.global.gpu.add.u32 [%0], 1;"
+                   :
+                   : "l"(counter)
+                   : "memory");
+  #else
+      atomicAdd(counter, 1);
+  #endif
+    }
+  } else {
+    asm volatile("barrier.cta.arrive 1, %0;" : : "r"(blockDim.x) : "memory");
+  }
+#else
+  __syncthreads();
+  if (threadIdx.x == 0) atomicAdd(counter, 1);
+#endif
+}
+
+template <typename T, int ngpus>
+__global__ void __launch_bounds__(kMnnvlLamportAgThreads, 1)
+    mnnvl_lamport_all_gather(RankData* _dp, const T* __restrict__ input,
+                             T* __restrict__ result,
+                             T* __restrict__ multicast_buffer,
+                             uint32_t* __restrict__ epochs, int rank,
+                             int size_per_rank, int stage_size) {
+  using P = typename packed_t<T>::P;
+#if !defined(USE_ROCM) && CUDA_VERSION >= 12000 && defined(__CUDA_ARCH__) && \
+    (__CUDA_ARCH__ >= 900)
+  cudaGridDependencySynchronize();
+#endif
+  auto dp = *_dp;
+  int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  int stride = gridDim.x * blockDim.x;
+  uint32_t epoch = epochs[0];
+  int current_stage = epoch % 3;
+  int dirty_stage = (epoch + 1) % 3;
+  int dirty_size = epochs[2 + dirty_stage];
+  auto local_buffer = reinterpret_cast<P*>(const_cast<void*>(dp.ptrs[rank]));
+  auto current_local = local_buffer + current_stage * stage_size;
+  auto dirty_local = local_buffer + dirty_stage * stage_size;
+  auto current_multicast =
+      reinterpret_cast<P*>(multicast_buffer) + current_stage * stage_size;
+  auto packed_input = reinterpret_cast<const P*>(input);
+  auto packed_result = reinterpret_cast<P*>(result);
+
+  int total_size = size_per_rank * ngpus;
+  P local_value;
+  if (tid < size_per_rank) {
+    local_value = packed_input[tid];
+    current_multicast[rank * size_per_rank + tid] =
+        sanitize_lamport_payload(local_value);
+  }
+#if !defined(USE_ROCM) && CUDA_VERSION >= 12000 && defined(__CUDA_ARCH__) && \
+    (__CUDA_ARCH__ >= 900)
+  cudaTriggerProgrammaticLaunchCompletion();
+#endif
+
+  lamport_cta_arrive(&epochs[1]);
+
+  for (int idx = tid; idx < dirty_size; idx += stride) {
+    dirty_local[idx] = lamport_sentinel<P>();
+  }
+
+  if (tid < size_per_rank) {
+#pragma unroll
+    for (int src_rank = 0; src_rank < ngpus; ++src_rank) {
+      int output_idx = src_rank * size_per_rank + tid;
+      P value = src_rank == rank
+                    ? local_value
+                    : wait_lamport_payload(current_local + output_idx);
+      packed_result[output_idx] = value;
+    }
+  }
+
+  if (tid == 0) {
+    while (*reinterpret_cast<volatile uint32_t*>(&epochs[1]) < gridDim.x);
+    epochs[2 + current_stage] = total_size;
+    epochs[0] = epoch + 1;
+    epochs[1] = 0;
+  }
+}
+
+template <typename T, int ngpus>
+__global__ void __launch_bounds__(kMnnvlLamportRsThreads, 1)
+    mnnvl_lamport_reduce_scatter_kernel(RankData* _dp,
+                                        const T* __restrict__ input,
+                                        T* __restrict__ result,
+                                        uint32_t* __restrict__ epochs, int rank,
+                                        int size_per_rank, int stage_size) {
+  using P = typename packed_t<T>::P;
+  using A = typename packed_t<T>::A;
+#if !defined(USE_ROCM) && CUDA_VERSION >= 12000 && defined(__CUDA_ARCH__) && \
+    (__CUDA_ARCH__ >= 900)
+  cudaGridDependencySynchronize();
+#endif
+  auto dp = *_dp;
+  int dst_rank = blockIdx.x % ngpus;
+  int tile = blockIdx.x / ngpus;
+  int idx = tile * blockDim.x + threadIdx.x;
+  int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  int stride = gridDim.x * blockDim.x;
+  uint32_t epoch = epochs[0];
+  int current_stage = epoch % 3;
+  int dirty_stage = (epoch + 1) % 3;
+  int dirty_size = epochs[2 + dirty_stage];
+  auto local_buffer = reinterpret_cast<P*>(const_cast<void*>(dp.ptrs[rank]));
+  auto current_local = local_buffer + current_stage * stage_size;
+  auto dirty_local = local_buffer + dirty_stage * stage_size;
+  auto packed_input = reinterpret_cast<const P*>(input);
+
+  if (idx < size_per_rank && dst_rank != rank) {
+    auto dst = reinterpret_cast<P*>(const_cast<void*>(dp.ptrs[dst_rank])) +
+               current_stage * stage_size + rank * size_per_rank;
+    auto src = packed_input + dst_rank * size_per_rank;
+    dst[idx] = sanitize_lamport_payload(src[idx]);
+  }
+#if !defined(USE_ROCM) && CUDA_VERSION >= 12000 && defined(__CUDA_ARCH__) && \
+    (__CUDA_ARCH__ >= 900)
+  cudaTriggerProgrammaticLaunchCompletion();
+#endif
+
+  lamport_cta_arrive(&epochs[1]);
+
+  for (int idx = tid; idx < dirty_size; idx += stride) {
+    dirty_local[idx] = lamport_sentinel<P>();
+  }
+
+  if (idx < size_per_rank && dst_rank == rank) {
+    if constexpr (ngpus == 4) {
+      if (size_per_rank > kMnnvlLamportConcurrentPollMaxPacks) {
+        reinterpret_cast<P*>(result)[idx] =
+            reduce_lamport_payloads<P, A, ngpus>(current_local, packed_input,
+                                                 rank, size_per_rank, idx);
+      } else {
+        P values[ngpus];
+        wait_lamport_payloads<P, ngpus>(
+            current_local + idx, rank, size_per_rank,
+            packed_input[rank * size_per_rank + idx], values);
+        A tmp = upcast(values[0]);
+#pragma unroll
+        for (int src_rank = 1; src_rank < ngpus; ++src_rank) {
+          packed_assign_add(tmp, upcast(values[src_rank]));
+        }
+        reinterpret_cast<P*>(result)[idx] =
+            sanitize_lamport_payload(downcast<P>(tmp));
+      }
+    } else {
+      reinterpret_cast<P*>(result)[idx] = reduce_lamport_payloads<P, A, ngpus>(
+          current_local, packed_input, rank, size_per_rank, idx);
+    }
+  }
+
+  if (tid == 0) {
+    while (*reinterpret_cast<volatile uint32_t*>(&epochs[1]) < gridDim.x);
+    epochs[2 + current_stage] = size_per_rank * ngpus;
+    epochs[0] = epoch + 1;
+    epochs[1] = 0;
+  }
+}
+
+}  // namespace vllm

csrc/custom_all_reduce.cuh

@@ -1,299 +1,8 @@
 #pragma once
 
-#include <cuda.h>
-#include <cuda_bf16.h>
-#include <cuda_fp16.h>
-#include <cuda_runtime.h>
-
-#if defined(USE_ROCM)
-typedef __hip_bfloat16 nv_bfloat16;
-#endif
-
-#include <iostream>
-#include <array>
-#include <limits>
-#include <map>
-#include <unordered_map>
-#include <vector>
-#include <cstdlib>
-#include <cstring>
+#include "custom_collective_common.cuh"
 
 namespace vllm {
-#define CUDACHECK(cmd)                                              \
-  do {                                                              \
-    cudaError_t e = cmd;                                            \
-    if (e != cudaSuccess) {                                         \
-      printf("Failed: Cuda error %s:%d '%s'\n", __FILE__, __LINE__, \
-             cudaGetErrorString(e));                                \
-      exit(EXIT_FAILURE);                                           \
-    }                                                               \
-  } while (0)
-
-// Maximal number of blocks in allreduce kernel.
-constexpr int kMaxBlocks = 36;
-
-// Default number of blocks in allreduce kernel.
-#ifndef USE_ROCM
-const int defaultBlockLimit = 36;
-CUpointer_attribute rangeStartAddrAttr = CU_POINTER_ATTRIBUTE_RANGE_START_ADDR;
-#else
-const int defaultBlockLimit = 16;
-hipPointer_attribute rangeStartAddrAttr =
-    HIP_POINTER_ATTRIBUTE_RANGE_START_ADDR;
-#endif
-
-// Counter may overflow, but it's fine since unsigned int overflow is
-// well-defined behavior.
-using FlagType = uint32_t;
-
-// Two sets of peer counters are needed for two syncs: starting and ending an
-// operation. The reason is that it's possible for peer GPU block to arrive at
-// the second sync point while the current GPU block haven't passed the first
-// sync point. Thus, peer GPU may write counter+1 while current GPU is busy
-// waiting for counter. We use alternating counter array to avoid this
-// possibility.
-struct Signal {
-  alignas(128) FlagType start[kMaxBlocks][8];
-  alignas(128) FlagType end[kMaxBlocks][8];
-  alignas(128) FlagType _flag[kMaxBlocks];  // incremental flags for each rank
-};
-
-struct __align__(16) RankData {
-  const void* ptrs[8];
-};
-
-struct __align__(16) RankSignals {
-  Signal* signals[8];
-};
-
-// like std::array, but aligned
-template <typename T, int sz>
-struct __align__(alignof(T) * sz) array_t {
-  T data[sz];
-  using type = T;
-  static constexpr int size = sz;
-};
-
-// use packed type to maximize memory efficiency
-// goal: generate ld.128 and st.128 instructions
-template <typename T>
-struct packed_t {
-  // the (P)acked type for load/store
-  using P = array_t<T, 16 / sizeof(T)>;
-  // the (A)ccumulator type for reduction
-  using A = array_t<float, 16 / sizeof(T)>;
-};
-
-#define DINLINE __device__ __forceinline__
-
-// scalar cast functions
-DINLINE float upcast_s(half val) { return __half2float(val); }
-
-template <typename T>
-DINLINE T downcast_s(float val);
-template <>
-DINLINE half downcast_s(float val) {
-  return __float2half(val);
-}
-
-// scalar add functions
-// for some reason when compiling with Pytorch, the + operator for half and
-// bfloat is disabled so we call the intrinsics directly
-DINLINE half& assign_add(half& a, half b) {
-  a = __hadd(a, b);
-  return a;
-}
-DINLINE float& assign_add(float& a, float b) { return a += b; }
-
-#if (__CUDA_ARCH__ >= 800 || !defined(__CUDA_ARCH__))
-DINLINE float upcast_s(nv_bfloat16 val) { return __bfloat162float(val); }
-template <>
-DINLINE nv_bfloat16 downcast_s(float val) {
-  return __float2bfloat16(val);
-}
-DINLINE nv_bfloat16& assign_add(nv_bfloat16& a, nv_bfloat16 b) {
-  a = __hadd(a, b);
-  return a;
-}
-#endif
-
-template <typename T, int N>
-DINLINE array_t<T, N>& packed_assign_add(array_t<T, N>& a, array_t<T, N> b) {
-#pragma unroll
-  for (int i = 0; i < N; i++) {
-    assign_add(a.data[i], b.data[i]);
-  }
-  return a;
-}
-
-template <typename T, int N>
-DINLINE array_t<float, N> upcast(array_t<T, N> val) {
-  if constexpr (std::is_same<T, float>::value) {
-    return val;
-  } else {
-    array_t<float, N> out;
-#pragma unroll
-    for (int i = 0; i < N; i++) {
-      out.data[i] = upcast_s(val.data[i]);
-    }
-    return out;
-  }
-}
-
-template <typename O>
-DINLINE O downcast(array_t<float, O::size> val) {
-  if constexpr (std::is_same<typename O::type, float>::value) {
-    return val;
-  } else {
-    O out;
-#pragma unroll
-    for (int i = 0; i < O::size; i++) {
-      out.data[i] = downcast_s<typename O::type>(val.data[i]);
-    }
-    return out;
-  }
-}
-
-#if !defined(USE_ROCM)
-
-static DINLINE void st_flag_release(FlagType* flag_addr, FlagType flag) {
-  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
-  asm volatile("st.release.sys.global.u32 [%1], %0;" ::"r"(flag),
-               "l"(flag_addr));
-  #else
-  asm volatile("membar.sys; st.volatile.global.u32 [%1], %0;" ::"r"(flag),
-               "l"(flag_addr));
-  #endif
-}
-
-static DINLINE FlagType ld_flag_acquire(FlagType* flag_addr) {
-  FlagType flag;
-  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
-  asm volatile("ld.acquire.sys.global.u32 %0, [%1];"
-               : "=r"(flag)
-               : "l"(flag_addr));
-  #else
-  asm volatile("ld.volatile.global.u32 %0, [%1]; membar.gl;"
-               : "=r"(flag)
-               : "l"(flag_addr));
-  #endif
-  return flag;
-}
-
-static DINLINE void st_flag_volatile(FlagType* flag_addr, FlagType flag) {
-  asm volatile("st.volatile.global.u32 [%1], %0;" ::"r"(flag), "l"(flag_addr));
-}
-
-static DINLINE FlagType ld_flag_volatile(FlagType* flag_addr) {
-  FlagType flag;
-  asm volatile("ld.volatile.global.u32 %0, [%1];"
-               : "=r"(flag)
-               : "l"(flag_addr));
-  return flag;
-}
-
-// This function is meant to be used as the first synchronization in the all
-// reduce kernel. Thus, it doesn't need to make any visibility guarantees for
-// prior memory accesses. Note: volatile writes will not be reordered against
-// other volatile writes.
-template <int ngpus>
-DINLINE void barrier_at_start(const RankSignals& sg, Signal* self_sg,
-                              int rank) {
-  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
-  if (threadIdx.x < ngpus) {
-    auto peer_counter_ptr = &sg.signals[threadIdx.x]->start[blockIdx.x][rank];
-    auto self_counter_ptr = &self_sg->start[blockIdx.x][threadIdx.x];
-    // Write the expected counter value to peer and wait for correct value
-    // from peer.
-    st_flag_volatile(peer_counter_ptr, flag);
-    while (ld_flag_volatile(self_counter_ptr) != flag);
-  }
-  __syncthreads();
-  // use one thread to update flag
-  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
-}
-
-// This function is meant to be used as the second or the final
-// synchronization barrier in the all reduce kernel. If it's the final
-// synchronization barrier, we don't need to make any visibility guarantees
-// for prior memory accesses.
-template <int ngpus, bool final_sync = false>
-DINLINE void barrier_at_end(const RankSignals& sg, Signal* self_sg, int rank) {
-  __syncthreads();
-  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
-  if (threadIdx.x < ngpus) {
-    auto peer_counter_ptr = &sg.signals[threadIdx.x]->end[blockIdx.x][rank];
-    auto self_counter_ptr = &self_sg->end[blockIdx.x][threadIdx.x];
-    // Write the expected counter value to peer and wait for correct value from
-    // peer.
-    if constexpr (!final_sync) {
-      st_flag_release(peer_counter_ptr, flag);
-      while (ld_flag_acquire(self_counter_ptr) != flag);
-    } else {
-      st_flag_volatile(peer_counter_ptr, flag);
-      while (ld_flag_volatile(self_counter_ptr) != flag);
-    }
-  }
-  if constexpr (!final_sync) __syncthreads();
-
-  // use one thread to update flag
-  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
-}
-
-#else
-
-template <int ngpus>
-DINLINE void barrier_at_start(const RankSignals& sg, Signal* self_sg,
-                              int rank) {
-  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
-  if (threadIdx.x < ngpus) {
-    // simultaneously write to the corresponding flag of all ranks.
-    // Latency = 1 p2p write
-    __scoped_atomic_store_n(&sg.signals[threadIdx.x]->start[blockIdx.x][rank],
-                            flag, __ATOMIC_RELAXED, __MEMORY_SCOPE_SYSTEM);
-    // wait until we got true from all ranks
-    while (__scoped_atomic_load_n(&self_sg->start[blockIdx.x][threadIdx.x],
-                                  __ATOMIC_RELAXED,
-                                  __MEMORY_SCOPE_DEVICE) < flag);
-  }
-  __syncthreads();
-  // use one thread to update flag
-  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
-}
-
-template <int ngpus, bool final_sync = false>
-DINLINE void barrier_at_end(const RankSignals& sg, Signal* self_sg, int rank) {
-  __syncthreads();
-  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
-  if (threadIdx.x < ngpus) {
-    // simultaneously write to the corresponding flag of all ranks.
-    // Latency = 1 p2p write
-    __scoped_atomic_store_n(&sg.signals[threadIdx.x]->end[blockIdx.x][rank],
-                            flag,
-                            final_sync ? __ATOMIC_RELAXED : __ATOMIC_RELEASE,
-                            __MEMORY_SCOPE_SYSTEM);
-    // wait until we got true from all ranks
-    while (
-        __scoped_atomic_load_n(&self_sg->end[blockIdx.x][threadIdx.x],
-                               final_sync ? __ATOMIC_RELAXED : __ATOMIC_ACQUIRE,
-                               __MEMORY_SCOPE_DEVICE) < flag);
-  }
-  if constexpr (!final_sync) __syncthreads();
-  // use one thread to update flag
-  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
-}
-
-#endif
-
-template <typename P, int ngpus, typename A>
-DINLINE P packed_reduce(const P* ptrs[], int idx) {
-  A tmp = upcast(ptrs[0][idx]);
-#pragma unroll
-  for (int i = 1; i < ngpus; i++) {
-    packed_assign_add(tmp, upcast(ptrs[i][idx]));
-  }
-  return downcast<P>(tmp);
-}
 
 template <typename T, int ngpus>
 __global__ void __launch_bounds__(512, 1)
@@ -616,6 +325,21 @@ class CustomAllreduce {
 #undef KL
   }
 
+  void allgather(cudaStream_t stream, void* input, void* output, int size_bytes,
+                 int threads = 512, int block_limit = defaultBlockLimit);
+  template <typename T>
+  void mnnvl_lamport_allgather(cudaStream_t stream, T* input, T* output,
+                               void* local_buffer, void* multicast_buffer,
+                               uint32_t* epochs, int size_bytes,
+                               int stage_size_bytes);
+  template <typename T>
+  void reduce_scatter(cudaStream_t stream, T* input, T* output, int size,
+                      int threads = 512, int block_limit = defaultBlockLimit);
+  template <typename T>
+  void mnnvl_lamport_reduce_scatter(cudaStream_t stream, T* input, T* output,
+                                    void* local_buffer, uint32_t* epochs,
+                                    int size, int stage_size_bytes);
+
   ~CustomAllreduce() {
     for (auto [_, ptr] : ipc_handles_) {
       CUDACHECK(cudaIpcCloseMemHandle(ptr));
@@ -625,8 +349,8 @@ class CustomAllreduce {
 
 /**
  * To inspect PTX/SASS, copy paste this header file to compiler explorer and
- add a template instantiation:
+ * add a template instantiation:
  * template void vllm::CustomAllreduce::allreduce<half>(cudaStream_t, half *,
- half *, int, int, int);
-*/
-}  // namespace vllm
+ *                                                       half *, int, int, int);
+ */
+}  // namespace vllm

csrc/custom_collective_common.cuh

@@ -0,0 +1,332 @@
+#pragma once
+
+#include <cuda.h>
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+
+#if defined(USE_ROCM)
+typedef __hip_bfloat16 nv_bfloat16;
+#endif
+
+#include <iostream>
+#include <array>
+#include <limits>
+#include <map>
+#include <unordered_map>
+#include <vector>
+#include <cstdlib>
+#include <cstring>
+
+namespace vllm {
+constexpr int kMaxCustomCollectiveRanks = 16;
+
+#define CUDACHECK(cmd)                                              \
+  do {                                                              \
+    cudaError_t e = cmd;                                            \
+    if (e != cudaSuccess) {                                         \
+      printf("Failed: Cuda error %s:%d '%s'\n", __FILE__, __LINE__, \
+             cudaGetErrorString(e));                                \
+      exit(EXIT_FAILURE);                                           \
+    }                                                               \
+  } while (0)
+
+// Maximal number of blocks in allreduce kernel.
+constexpr int kMaxBlocks = 36;
+
+// Default number of blocks in allreduce kernel.
+#ifndef USE_ROCM
+inline constexpr int defaultBlockLimit = 36;
+inline CUpointer_attribute rangeStartAddrAttr =
+    CU_POINTER_ATTRIBUTE_RANGE_START_ADDR;
+#else
+inline constexpr int defaultBlockLimit = 16;
+inline hipPointer_attribute rangeStartAddrAttr =
+    HIP_POINTER_ATTRIBUTE_RANGE_START_ADDR;
+#endif
+
+// Counter may overflow, but it's fine since unsigned int overflow is
+// well-defined behavior.
+using FlagType = uint32_t;
+
+// Two sets of peer counters are needed for two syncs: starting and ending an
+// operation. The reason is that it's possible for peer GPU block to arrive at
+// the second sync point while the current GPU block haven't passed the first
+// sync point. Thus, peer GPU may write counter+1 while current GPU is busy
+// waiting for counter. We use alternating counter array to avoid this
+// possibility.
+struct Signal {
+  alignas(128) FlagType start[kMaxBlocks][kMaxCustomCollectiveRanks];
+  alignas(128) FlagType end[kMaxBlocks][kMaxCustomCollectiveRanks];
+  alignas(128) FlagType _flag[kMaxBlocks];  // incremental flags for each rank
+};
+
+struct __align__(16) RankData {
+  const void* ptrs[kMaxCustomCollectiveRanks];
+};
+
+struct __align__(16) RankSignals {
+  Signal* signals[kMaxCustomCollectiveRanks];
+};
+
+// like std::array, but aligned
+template <typename T, int sz>
+struct __align__(alignof(T) * sz) array_t {
+  T data[sz];
+  using type = T;
+  static constexpr int size = sz;
+};
+
+// use packed type to maximize memory efficiency
+// goal: generate ld.128 and st.128 instructions
+template <typename T>
+struct packed_t {
+  // the (P)acked type for load/store
+  using P = array_t<T, 16 / sizeof(T)>;
+  // the (A)ccumulator type for reduction
+  using A = array_t<float, 16 / sizeof(T)>;
+};
+
+#define DINLINE __device__ __forceinline__
+
+// scalar cast functions
+DINLINE float upcast_s(half val) { return __half2float(val); }
+
+template <typename T>
+DINLINE T downcast_s(float val);
+template <>
+DINLINE half downcast_s(float val) {
+  return __float2half(val);
+}
+
+// scalar add functions
+// for some reason when compiling with Pytorch, the + operator for half and
+// bfloat is disabled so we call the intrinsics directly
+DINLINE half& assign_add(half& a, half b) {
+  a = __hadd(a, b);
+  return a;
+}
+DINLINE float& assign_add(float& a, float b) { return a += b; }
+
+#if (__CUDA_ARCH__ >= 800 || !defined(__CUDA_ARCH__))
+DINLINE float upcast_s(nv_bfloat16 val) { return __bfloat162float(val); }
+template <>
+DINLINE nv_bfloat16 downcast_s(float val) {
+  return __float2bfloat16(val);
+}
+DINLINE nv_bfloat16& assign_add(nv_bfloat16& a, nv_bfloat16 b) {
+  a = __hadd(a, b);
+  return a;
+}
+#endif
+
+template <typename T, int N>
+DINLINE array_t<T, N>& packed_assign_add(array_t<T, N>& a, array_t<T, N> b) {
+#pragma unroll
+  for (int i = 0; i < N; i++) {
+    assign_add(a.data[i], b.data[i]);
+  }
+  return a;
+}
+
+template <typename T, int N>
+DINLINE array_t<float, N> upcast(array_t<T, N> val) {
+  if constexpr (std::is_same<T, float>::value) {
+    return val;
+  } else {
+    array_t<float, N> out;
+#pragma unroll
+    for (int i = 0; i < N; i++) {
+      out.data[i] = upcast_s(val.data[i]);
+    }
+    return out;
+  }
+}
+
+template <typename O>
+DINLINE O downcast(array_t<float, O::size> val) {
+  if constexpr (std::is_same<typename O::type, float>::value) {
+    return val;
+  } else {
+    O out;
+#pragma unroll
+    for (int i = 0; i < O::size; i++) {
+      out.data[i] = downcast_s<typename O::type>(val.data[i]);
+    }
+    return out;
+  }
+}
+
+#if !defined(USE_ROCM)
+
+static DINLINE void st_flag_release(FlagType* flag_addr, FlagType flag) {
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  asm volatile("st.release.sys.global.u32 [%1], %0;" ::"r"(flag),
+               "l"(flag_addr));
+  #else
+  asm volatile("membar.sys; st.volatile.global.u32 [%1], %0;" ::"r"(flag),
+               "l"(flag_addr));
+  #endif
+}
+
+static DINLINE FlagType ld_flag_acquire(FlagType* flag_addr) {
+  FlagType flag;
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
+  asm volatile("ld.acquire.sys.global.u32 %0, [%1];"
+               : "=r"(flag)
+               : "l"(flag_addr));
+  #else
+  asm volatile("ld.volatile.global.u32 %0, [%1]; membar.gl;"
+               : "=r"(flag)
+               : "l"(flag_addr));
+  #endif
+  return flag;
+}
+
+static DINLINE void st_flag_volatile(FlagType* flag_addr, FlagType flag) {
+  asm volatile("st.volatile.global.u32 [%1], %0;" ::"r"(flag), "l"(flag_addr));
+}
+
+static DINLINE FlagType ld_flag_volatile(FlagType* flag_addr) {
+  FlagType flag;
+  asm volatile("ld.volatile.global.u32 %0, [%1];"
+               : "=r"(flag)
+               : "l"(flag_addr));
+  return flag;
+}
+
+// This function is meant to be used as the first synchronization in the all
+// reduce kernel. Thus, it doesn't need to make any visibility guarantees for
+// prior memory accesses. Note: volatile writes will not be reordered against
+// other volatile writes.
+template <int ngpus>
+DINLINE void barrier_at_start(const RankSignals& sg, Signal* self_sg,
+                              int rank) {
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    auto peer_counter_ptr = &sg.signals[threadIdx.x]->start[blockIdx.x][rank];
+    auto self_counter_ptr = &self_sg->start[blockIdx.x][threadIdx.x];
+    // Write the expected counter value to peer and wait for correct value
+    // from peer.
+    st_flag_volatile(peer_counter_ptr, flag);
+    while (ld_flag_volatile(self_counter_ptr) != flag);
+  }
+  __syncthreads();
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+template <int ngpus>
+DINLINE void barrier_at_start_release(const RankSignals& sg, Signal* self_sg,
+                                      int rank) {
+  __syncthreads();
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    auto peer_counter_ptr = &sg.signals[threadIdx.x]->start[blockIdx.x][rank];
+    auto self_counter_ptr = &self_sg->start[blockIdx.x][threadIdx.x];
+    st_flag_release(peer_counter_ptr, flag);
+    while (ld_flag_acquire(self_counter_ptr) != flag);
+  }
+  __syncthreads();
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+// This function is meant to be used as the second or the final
+// synchronization barrier in the all reduce kernel. If it's the final
+// synchronization barrier, we don't need to make any visibility guarantees
+// for prior memory accesses.
+template <int ngpus, bool final_sync = false>
+DINLINE void barrier_at_end(const RankSignals& sg, Signal* self_sg, int rank) {
+  __syncthreads();
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    auto peer_counter_ptr = &sg.signals[threadIdx.x]->end[blockIdx.x][rank];
+    auto self_counter_ptr = &self_sg->end[blockIdx.x][threadIdx.x];
+    // Write the expected counter value to peer and wait for correct value from
+    // peer.
+    if constexpr (!final_sync) {
+      st_flag_release(peer_counter_ptr, flag);
+      while (ld_flag_acquire(self_counter_ptr) != flag);
+    } else {
+      st_flag_volatile(peer_counter_ptr, flag);
+      while (ld_flag_volatile(self_counter_ptr) != flag);
+    }
+  }
+  if constexpr (!final_sync) __syncthreads();
+
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+#else
+
+template <int ngpus>
+DINLINE void barrier_at_start(const RankSignals& sg, Signal* self_sg,
+                              int rank) {
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    // simultaneously write to the corresponding flag of all ranks.
+    // Latency = 1 p2p write
+    __scoped_atomic_store_n(&sg.signals[threadIdx.x]->start[blockIdx.x][rank],
+                            flag, __ATOMIC_RELAXED, __MEMORY_SCOPE_SYSTEM);
+    // wait until we got true from all ranks
+    while (__scoped_atomic_load_n(&self_sg->start[blockIdx.x][threadIdx.x],
+                                  __ATOMIC_RELAXED,
+                                  __MEMORY_SCOPE_DEVICE) < flag);
+  }
+  __syncthreads();
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+template <int ngpus>
+DINLINE void barrier_at_start_release(const RankSignals& sg, Signal* self_sg,
+                                      int rank) {
+  __syncthreads();
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    __scoped_atomic_store_n(&sg.signals[threadIdx.x]->start[blockIdx.x][rank],
+                            flag, __ATOMIC_RELEASE, __MEMORY_SCOPE_SYSTEM);
+    while (__scoped_atomic_load_n(&self_sg->start[blockIdx.x][threadIdx.x],
+                                  __ATOMIC_ACQUIRE,
+                                  __MEMORY_SCOPE_DEVICE) < flag);
+  }
+  __syncthreads();
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+template <int ngpus, bool final_sync = false>
+DINLINE void barrier_at_end(const RankSignals& sg, Signal* self_sg, int rank) {
+  __syncthreads();
+  uint32_t flag = self_sg->_flag[blockIdx.x] + 1;
+  if (threadIdx.x < ngpus) {
+    // simultaneously write to the corresponding flag of all ranks.
+    // Latency = 1 p2p write
+    __scoped_atomic_store_n(&sg.signals[threadIdx.x]->end[blockIdx.x][rank],
+                            flag,
+                            final_sync ? __ATOMIC_RELAXED : __ATOMIC_RELEASE,
+                            __MEMORY_SCOPE_SYSTEM);
+    // wait until we got true from all ranks
+    while (
+        __scoped_atomic_load_n(&self_sg->end[blockIdx.x][threadIdx.x],
+                               final_sync ? __ATOMIC_RELAXED : __ATOMIC_ACQUIRE,
+                               __MEMORY_SCOPE_DEVICE) < flag);
+  }
+  if constexpr (!final_sync) __syncthreads();
+  // use one thread to update flag
+  if (threadIdx.x == 0) self_sg->_flag[blockIdx.x] = flag;
+}
+
+#endif
+
+template <typename P, int ngpus, typename A>
+DINLINE P packed_reduce(const P* ptrs[], int idx) {
+  A tmp = upcast(ptrs[0][idx]);
+#pragma unroll
+  for (int i = 1; i < ngpus; i++) {
+    packed_assign_add(tmp, upcast(ptrs[i][idx]));
+  }
+  return downcast<P>(tmp);
+}
+
+}  // namespace vllm