[Rust Frontend] Let the server run over any Llm

.buildkite/hardware_tests/cpu.yaml

@@ -257,8 +257,11 @@ steps:
   agents:
     queue: amd-cpu
   source_file_dependencies:
+    - setup.py
     - .buildkite/hardware_tests/cpu.yaml
+    - .buildkite/hardware_tests/zen.yaml
     - .buildkite/image_build/image_build_zen_cpu.sh
+    - .buildkite/scripts/hardware_ci/run-zen-cpu-test.sh
     - docker/Dockerfile.zen
     - docker/Dockerfile.cpu
     - requirements/cpu.txt

.buildkite/hardware_tests/zen.yaml

@@ -0,0 +1,107 @@
+group: Zen CPU
+depends_on: []
+steps:
+- label: Zen-CPU-Kernel Tests
+  key: zen-cpu-kernel-tests
+  # run-zen-cpu-test.sh pulls the prebuilt <repo>:<commit>-zen-cpu image
+  # that zen-cpu-image-build publishes, falling back to a local build
+  # only if the pull misses.
+  depends_on:
+  - zen-cpu-image-build
+  allow_dependency_failure: true
+  soft_fail: true
+  device: zen5
+  no_plugin: true
+  source_file_dependencies:
+  - setup.py
+  - .buildkite/hardware_tests/cpu.yaml
+  - .buildkite/hardware_tests/zen.yaml
+  - .buildkite/image_build/image_build_zen_cpu.sh
+  - .buildkite/scripts/hardware_ci/run-zen-cpu-test.sh
+  - docker/Dockerfile.zen
+  - docker/Dockerfile.cpu
+  - requirements/cpu.txt
+  - requirements/build/cpu.txt
+  - requirements/test/cpu.txt
+  - cmake/cpu_extension.cmake
+  - vllm/model_executor/layers/utils.py
+  - vllm/model_executor/kernels/linear/zentorch_utils.py
+  - vllm/model_executor/kernels/linear/scaled_mm/zentorch.py
+  - vllm/model_executor/kernels/linear/mixed_precision/zentorch.py
+  - vllm/platforms/cpu.py
+  - vllm/platforms/zen_cpu.py
+  - vllm/platforms/__init__.py
+  - tests/model_executor/test_cpu_unquantized_gemm_dispatch.py
+  - tests/test_zen_cpu_platform_detection.py
+  - tests/kernels/moe/test_zen_cpu_fused_moe.py
+  - tests/kernels/moe/test_zen_cpu_int8_moe.py
+  - vllm/v1/attention/backends/zentorch_sdpa.py
+  - tests/test_zentorch_version_pin.py
+  - vllm/v1/attention/backends/mla/prefill/zen_cpu_sdpa.py
+  - vllm/model_executor/layers/fused_moe/experts/zentorch_moe.py
+  - tests/kernels/quantization/test_zen_da8w4.py
+  commands:
+    - |
+      # The zen5 agent env does not provide REGISTRY/REPO, so pass the public-ECR
+      # coordinates the zen-cpu-image-build step pushed to (anonymous pull). Keep
+      # the repo mapping in sync with ci_config.yaml (main vs premerge).
+      export REGISTRY="public.ecr.aws/q9t5s3a7"
+      if [[ "$$BUILDKITE_BRANCH" == "main" ]]; then
+        export REPO="vllm-ci-postmerge-repo"
+      else
+        export REPO="vllm-ci-test-repo"
+      fi
+      bash .buildkite/scripts/hardware_ci/run-zen-cpu-test.sh 20m "
+      pytest -x -v -s tests/model_executor/test_cpu_unquantized_gemm_dispatch.py"
+
+- label: Zen-CPU-Platform-Detection Tests
+  key: zen-cpu-platform-detection-tests
+  # Same model as the kernel step: run-zen-cpu-test.sh pulls the prebuilt
+  # <repo>:<commit>-zen-cpu image from zen-cpu-image-build (local-build fallback).
+  depends_on:
+  - zen-cpu-image-build
+  allow_dependency_failure: true
+  soft_fail: true
+  device: zen5
+  no_plugin: true
+  source_file_dependencies:
+  - setup.py
+  - .buildkite/hardware_tests/cpu.yaml
+  - .buildkite/hardware_tests/zen.yaml
+  - .buildkite/image_build/image_build_zen_cpu.sh
+  - .buildkite/scripts/hardware_ci/run-zen-cpu-test.sh
+  - docker/Dockerfile.zen
+  - docker/Dockerfile.cpu
+  - requirements/cpu.txt
+  - requirements/build/cpu.txt
+  - requirements/test/cpu.txt
+  - cmake/cpu_extension.cmake
+  - vllm/model_executor/layers/utils.py
+  - vllm/model_executor/kernels/linear/zentorch_utils.py
+  - vllm/model_executor/kernels/linear/scaled_mm/zentorch.py
+  - vllm/model_executor/kernels/linear/mixed_precision/zentorch.py
+  - vllm/platforms/cpu.py
+  - vllm/platforms/zen_cpu.py
+  - vllm/platforms/__init__.py
+  - tests/model_executor/test_cpu_unquantized_gemm_dispatch.py
+  - tests/test_zen_cpu_platform_detection.py
+  - tests/kernels/moe/test_zen_cpu_fused_moe.py
+  - tests/kernels/moe/test_zen_cpu_int8_moe.py
+  - vllm/v1/attention/backends/zentorch_sdpa.py
+  - tests/test_zentorch_version_pin.py
+  - vllm/v1/attention/backends/mla/prefill/zen_cpu_sdpa.py
+  - vllm/model_executor/layers/fused_moe/experts/zentorch_moe.py
+  - tests/kernels/quantization/test_zen_da8w4.py
+  commands:
+    - |
+      # The zen5 agent env does not provide REGISTRY/REPO, so pass the public-ECR
+      # coordinates the zen-cpu-image-build step pushed to (anonymous pull). Keep
+      # the repo mapping in sync with ci_config.yaml (main vs premerge).
+      export REGISTRY="public.ecr.aws/q9t5s3a7"
+      if [[ "$$BUILDKITE_BRANCH" == "main" ]]; then
+        export REPO="vllm-ci-postmerge-repo"
+      else
+        export REPO="vllm-ci-test-repo"
+      fi
+      bash .buildkite/scripts/hardware_ci/run-zen-cpu-test.sh 20m "
+      pytest -x -v -s tests/test_zen_cpu_platform_detection.py"

.buildkite/image_build/image_build_zen_cpu.sh

@@ -5,14 +5,6 @@
 # Build the AMD Zen CPU image (vLLM + zentorch) by layering docker/Dockerfile.zen
 # on top of a CPU base image that already has vLLM installed.
 #
-# To avoid recompiling vLLM from source, prefer reusing the shared CPU image that
-# the image-build-cpu step already builds and publishes for this commit
-# (`<repo>:<commit>-cpu`). It lives in public ECR, which allows anonymous pulls,
-# so no registry credentials are required. If the pull fails (e.g. the image
-# isn't published yet), fall back to building the base from docker/Dockerfile.cpu.
-#
-# The zen image is not pushed to a registry.
-#
 # See docker/Dockerfile.zen for the build workflow this mirrors.
 set -euo pipefail
 
@@ -73,3 +65,14 @@ docker build --file docker/Dockerfile.zen \
   --tag "$IMAGE" \
   --target vllm-zen-test \
   --progress plain .
+
+# Step 3: push the zen image so test nodes can pull it instead of rebuilding.
+# Only when a registry is provided (CI); local runs keep the image local.
+# Pushing to public ECR requires AWS credentials on the build agent; pulls are
+# anonymous.
+if [[ -n "${REGISTRY}" ]]; then
+  echo "--- :docker: Pushing Zen image: $IMAGE"
+  aws ecr-public get-login-password --region us-east-1 \
+    | docker login --username AWS --password-stdin "$REGISTRY"
+  docker push "$IMAGE"
+fi

.buildkite/intel_jobs/entrypoints_intel.yaml

@@ -111,7 +111,9 @@ steps:
     bash .buildkite/scripts/hardware_ci/run-intel-test.sh
     'export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
     cd tests &&
-    pytest -v -s entrypoints/scale_out'
+    pytest -v -s entrypoints/scale_out
+    --deselect="entrypoints/scale_out/derender/test_derender_parity.py::test_inline_derender_parity[n2]"
+    --deselect="entrypoints/scale_out/derender/test_derender_parity.py::test_inline_derender_stream_parity[n2]"'
 
 - label: ":intel: (BMG) Entrypoints Integration (API Server Generate)"
   key: entrypoints-integration-api-server-generate-intel

.buildkite/intel_jobs/models_multimodal_intel.yaml

@@ -107,6 +107,7 @@ steps:
       --ignore models/multimodal/generation/test_whisper.py
       --ignore models/multimodal/generation/test_gemma_ragged_audio.py
       --ignore models/multimodal/generation/test_memory_leak.py
+      --ignore models/multimodal/generation/test_unlimited_ocr.py
       --ignore models/multimodal/processing'
 
 - label: ":intel: (BMG) Multi-Modal Processor" # 44min

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

@@ -0,0 +1,97 @@
+#!/bin/bash
+
+# Run the AMD Zen CPU tests on the zen5 hardware.
+
+set -euox pipefail
+
+# allow to bind to different cores
+CORE_RANGE=${CORE_RANGE:-0-47}
+NUMA_NODE=${NUMA_NODE:-0}
+IMAGE_NAME="zen-cpu-test-$NUMA_NODE"
+FALLBACK_BASE_IMAGE="zen-cpu-base-$NUMA_NODE"
+TIMEOUT_VAL=$1
+TEST_COMMAND=$2
+
+# Install zentorch directly in Dockerfile.zen instead of `uv pip install
+# "vllm[zen]"`, which makes uv re-resolve vLLM and fail on its transitive
+# triton-cpu URL dependency. Keep in sync with the `zen` extra in setup.py.
+ZENTORCH_VERSION="${ZENTORCH_VERSION:-2.14.0.1}"
+
+CLEANUP_IMAGES=()
+cleanup() {
+    if [ "${#CLEANUP_IMAGES[@]}" -gt 0 ]; then
+        docker rmi -f "${CLEANUP_IMAGES[@]}" > /dev/null 2>&1 || true
+    fi
+}
+trap cleanup EXIT
+
+# Published images (resolvable only when the Buildkite registry env vars are
+# present, i.e. in CI). Pulls from public ECR are anonymous.
+ZEN_IMAGE=""
+SHARED_CPU_IMAGE=""
+if [ -n "${REGISTRY:-}" ] && [ -n "${REPO:-}" ] && [ -n "${BUILDKITE_COMMIT:-}" ]; then
+    ZEN_IMAGE="$REGISTRY/$REPO:$BUILDKITE_COMMIT-zen-cpu"
+    SHARED_CPU_IMAGE="$REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu"
+fi
+
+if [ -n "$ZEN_IMAGE" ] && docker pull "$ZEN_IMAGE"; then
+    # Fast path: reuse the prebuilt zen image; no local build.
+    echo "--- :docker: Using published Zen image: $ZEN_IMAGE"
+    IMAGE_NAME="$ZEN_IMAGE"
+    CLEANUP_IMAGES+=("$ZEN_IMAGE")
+else
+    # Step 1: obtain the CPU base image that Dockerfile.zen layers on. Prefer
+    # pulling the published `-cpu` image; fall back to building it from source.
+    if [ -n "$SHARED_CPU_IMAGE" ] && docker pull "$SHARED_CPU_IMAGE"; then
+        echo "--- :docker: Using published CPU image as base: $SHARED_CPU_IMAGE"
+        BASE_IMAGE="$SHARED_CPU_IMAGE"
+        CLEANUP_IMAGES+=("$SHARED_CPU_IMAGE")
+    else
+        echo "--- :docker: Published CPU image unavailable; building base from source"
+        docker build --progress plain --tag "$FALLBACK_BASE_IMAGE" \
+            --target vllm-openai -f docker/Dockerfile.cpu .
+        BASE_IMAGE="$FALLBACK_BASE_IMAGE"
+        CLEANUP_IMAGES+=("$FALLBACK_BASE_IMAGE")
+    fi
+
+    # Step 2: build the zen test image on top of the CPU base.
+    echo "--- :docker: Building Zen test image"
+    docker build --progress plain --tag "$IMAGE_NAME" \
+        --build-arg BASE_IMAGE="$BASE_IMAGE" \
+        --build-arg ZENTORCH_VERSION="$ZENTORCH_VERSION" \
+        --target vllm-zen-test -f docker/Dockerfile.zen .
+    CLEANUP_IMAGES+=("$IMAGE_NAME")
+fi
+
+# Run the image, setting --shm-size=4g for tensor parallel. The in-container
+# command verifies the installed zentorch matches the pinned version and that
+# its torch.compile backend is registered (i.e. not the stable-only build), so
+# a mismatch fails loudly instead of silently testing a different zentorch.
+docker run --rm \
+    --cpuset-cpus="$CORE_RANGE" \
+    --cpuset-mems="$NUMA_NODE" \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    --privileged=true \
+    -e HF_TOKEN \
+    -e VLLM_CPU_KVCACHE_SPACE=16 \
+    -e VLLM_CPU_CI_ENV=1 \
+    -e VLLM_CPU_SIM_MULTI_NUMA=1 \
+    --shm-size=4g \
+    "$IMAGE_NAME" \
+    timeout "$TIMEOUT_VAL" bash -c "set -euox pipefail
+echo \"--- Print packages\"
+pip list
+echo \"--- Verify zentorch\"
+python -c \"
+import importlib.metadata as m, torch, zentorch
+from torch._dynamo import list_backends
+v = m.version('zentorch')
+assert v == '${ZENTORCH_VERSION}', (
+    'zentorch ' + v + ' != expected ${ZENTORCH_VERSION}'
+)
+assert 'zentorch' in list_backends(), (
+    'zentorch torch.compile backend not registered (stable-only build?)'
+)
+\"
+echo \"--- Running tests\"
+${TEST_COMMAND}"

.buildkite/test_areas/models_basic.yaml

@@ -407,6 +407,18 @@ steps:
       # PLE and QSA import conflicting HC registrations in a shared process.
       - pytest -v -s models/qwen4_exp/test_ple.py
 
+- label: ":nvidia: (B200) Qwen4 Exp HC Kernels"
+  key: qwen4-exp-hc-kernels-b200
+  timeout_in_minutes: 20
+  device: b200-k8s
+  source_file_dependencies:
+  - csrc/qwen4_exp/
+  - vllm/models/qwen4_exp/nvidia/hyperconnection.py
+  - vllm/models/qwen4_exp/nvidia/ops/hc.py
+  - tests/models/qwen4_exp/test_hc_ops.py
+  commands:
+    - pytest -v -s models/qwen4_exp/test_hc_ops.py
+
 - label: ":amd: (MI300) Qwen4 Exp"
   key: qwen4-exp-unit-tests-mi300
   automatic_shard: true

.buildkite/test_areas/models_multimodal.yaml

@@ -278,11 +278,11 @@ steps:
       depends_on:
       - image-build-amd
 
-- label: ":nvidia: (H200 MIG 18GB) [Nightly] Multimodal Models (Extended Pooling)"
+- label: ":nvidia: (H200 MIG 35GB) [Nightly] Multimodal Models (Extended Pooling)"
   key: multi-modal-models-extended-pooling
   automatic_shard: true
   optional: true
-  device: h200_18gb
+  device: h200_35gb
   source_file_dependencies:
   - vllm/
   - "!vllm/distributed/kv_transfer/"

CMakeLists.txt

@@ -1189,6 +1189,21 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
       "Building Kimi K3 AttnRes for archs: ${KIMI_K3_ATTN_RES_ARCHS}")
   endif()
 
+  # Qwen4Exp HyperConnection up projection + gate mix (tcgen05).
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(QWEN4_EXP_HC_UP_GATE_MIX_ARCHS
+      "10.0f" "${CUDA_ARCHS}")
+  endif()
+  if(QWEN4_EXP_HC_UP_GATE_MIX_ARCHS)
+    set(QWEN4_EXP_HC_UP_GATE_MIX_SRC "csrc/qwen4_exp/hc_up_gate_mix.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${QWEN4_EXP_HC_UP_GATE_MIX_SRC}"
+      CUDA_ARCHS "${QWEN4_EXP_HC_UP_GATE_MIX_ARCHS}")
+    list(APPEND VLLM_STABLE_EXT_SRC "${QWEN4_EXP_HC_UP_GATE_MIX_SRC}")
+    message(STATUS "Building Qwen4Exp HC up+gate mix for archs: "
+      "${QWEN4_EXP_HC_UP_GATE_MIX_ARCHS}")
+  endif()
+
   # Hadacore kernels
   cuda_archs_loose_intersection(HADACORE_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
   if(HADACORE_ARCHS)
@@ -1295,6 +1310,10 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
       target_compile_definitions(_C_stable_libtorch PRIVATE
         VLLM_ENABLE_KIMI_K3_ATTN_RES=1)
     endif()
+    if(QWEN4_EXP_HC_UP_GATE_MIX_ARCHS)
+      target_compile_definitions(_C_stable_libtorch PRIVATE
+        VLLM_ENABLE_QWEN4_EXP_HC_UP_GATE_MIX=1)
+    endif()
     if(BEAM_KERNEL_ARCHS)
       target_compile_definitions(_C_stable_libtorch PRIVATE
         VLLM_ENABLE_BEAM_KERNELS=1)

cmake/external_projects/deepgemm.cmake

@@ -38,8 +38,8 @@ else()
   # (vllm-project/DeepGEMM#43), plus the restored FP32 MQA logits and their
   # `logits_dtype` switch (vllm-project/DeepGEMM#46) and the mega_mhc
   # CUDA-graph replay barrier fix (vllm-project/DeepGEMM#48).
-  # Uses the PyTorch 2.11 stable ABI and CPython 3.10 limited API.
-  set(_DEEPGEMM_UPSTREAM_TAG "ae4fd25d036a86042ac4bbfc01e2e0fa5fb362b6")
+  # Targets the PyTorch 2.11 stable ABI (requires PyTorch 2.14+ headers at build time) and CPython 3.10 limited API.
+  set(_DEEPGEMM_UPSTREAM_TAG "82ca617e5879d0efb0aba5b91b1c81852588af6f")
 
   set(_deepgemm_fc_root "${FETCHCONTENT_BASE_DIR}")
   if(NOT _deepgemm_fc_root)