[MiniMax-M3] Retune decode indexer launch grid for occupancy (#130 )

Signed-off-by: Zijing Liu <liuzijing2014@gmail.com>
Add MiniMax-M3 modelopt nvfp4 support (#46380 )
2026-06-25 18:14:52 +00:00 · 2026-06-25 11:11:41 -07:00 · 2026-06-23 00:01:30 +00:00 · 2026-06-22 05:21:19 +00:00 · 2026-06-20 13:53:02 -07:00 · 2026-06-20 13:52:56 -07:00
163 changed files with 4358 additions and 1114 deletions
@@ -67,7 +67,7 @@ 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/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 &&
+        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:
@@ -647,7 +647,7 @@ steps:
  - pytest -v -s v1/cudagraph/test_cudagraph_mode.py

 - label: e2e Core (1 GPU) # TBD
-  timeout_in_minutes: 180
+  timeout_in_minutes: 35
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
  agent_pool: mi250_1
  optional: true
@@ -2075,19 +2075,6 @@ steps:
  - export VLLM_ALLOW_INSECURE_SERIALIZATION=1
  - pytest -v -s v1/spec_decode/test_acceptance_length.py -m slow_test

- label: e2e Core (1 GPU) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/v1/
-  - tests/v1/e2e/
-  - vllm/platforms/rocm.py
-  commands:
-  - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
-
 - label: e2e Scheduling (1 GPU) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
@@ -74,6 +74,16 @@ steps:
    - tests/v1/e2e/general/
  commands:
    - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
+  mirror:
+    amd:
+      device: mi250_1
+      timeout_in_minutes: 35
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/
+      - tests/v1/e2e/general/
+      - vllm/platforms/rocm.py

 - label: V1 e2e (2 GPUs)
  key: v1-e2e-2-gpus
@@ -104,6 +104,7 @@ steps:
      source_file_dependencies:
      - csrc/quantization/
      - vllm/model_executor/layers/quantization
+      - vllm/config/
      - tests/kernels/quantization
      - tests/kernels/quantization/test_rocm_skinny_gemms.py
      - vllm/_aiter_ops.py
@@ -101,6 +101,15 @@ steps:
  num_devices: 8
  commands:
    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200.txt
+  mirror:
+    amd:
+      device: mi300_8
+      timeout_in_minutes: 180
+      depends_on:
+      - image-build-amd
+      commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx.txt

 - label: MoE Refactor Integration Test (H100 - TEMPORARY)
  key: moe-refactor-integration-test-h100-temporary
@@ -68,7 +68,6 @@ 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:
-      soft_fail: true
      device: mi325_1
      depends_on:
      - image-build-amd
@@ -0,0 +1,35 @@
+---
+name: ci-fails-buildkite
+description: Fetch and diagnose vLLM Buildkite CI failure logs. Use when investigating failing CI jobs on a PR or build, when the user pastes a buildkite.com URL, or asks to fetch/diagnose CI logs.
+---
+
+# Diagnosing vLLM Buildkite CI Failures
+
+Buildkite logs are public; no login needed.
+
+`.buildkite/scripts/ci-fetch-log.sh` saves each log as `ci-<build>-<job-name>.log`, stripped of timestamps and ANSI codes. Existing files are kept; set `CI_FETCH_LOG_FORCE=1` to refetch.
+
+## Fetching logs
+
+```bash
+# All failed jobs in a PR's latest build (current branch's PR if omitted):
+.buildkite/scripts/ci-fetch-log.sh --pr <PR>
+
+# All failed jobs in a build (--soft also includes soft-failed jobs;
+# --all fetches every finished job):
+.buildkite/scripts/ci-fetch-log.sh "https://buildkite.com/vllm/ci/builds/<N>"
+
+# One job — `gh pr checks` URLs (#<job_uuid>) and web UI URLs (?sid=) both
+# work; pass "-" as a second argument to stream to stdout:
+.buildkite/scripts/ci-fetch-log.sh "https://buildkite.com/vllm/ci/builds/<N>#<job_uuid>"
+```
+
+To clean an already-downloaded log with `.buildkite/scripts/ci-clean-log.sh`:
+
+```bash
+./ci-clean-log.sh ci.log
+```
+
+## Reference
+
+See [docs/contributing/ci/failures.md](../../../docs/contributing/ci/failures.md) for the full guide: filing CI failure issues, investigating/bisecting, reproducing flaky tests, and daily triage.
@@ -2,15 +2,14 @@
 # for more info about CODEOWNERS file

 # This lists cover the "core" components of vLLM that require careful review
-/vllm/compilation @zou3519 @youkaichao @ProExpertProg @BoyuanFeng @vadiklyutiy
+/vllm/compilation @zou3519 @youkaichao @ProExpertProg @BoyuanFeng
 /vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery @xuechendi
 /vllm/lora @jeejeelee
 /vllm/model_executor/layers/attention @LucasWilkinson @MatthewBonanni
 /vllm/model_executor/layers/fused_moe @mgoin @pavanimajety @zyongye
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety @zyongye
 /vllm/model_executor/layers/mamba @tdoublep @tomeras91
-/vllm/model_executor/layers/mamba/gdn_linear_attn.py @tdoublep @ZJY0516 @vadiklyutiy
-/vllm/model_executor/layers/rotary_embedding.py @vadiklyutiy
+/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py @tdoublep @ZJY0516 @vadiklyutiy
 /vllm/model_executor/model_loader @22quinn
 /vllm/model_executor/layers/batch_invariant.py @yewentao256 
 /vllm/ir @ProExpertProg
@@ -199,7 +199,9 @@ cython_debug/
 .vscode/

 # Claude
-.claude/
+.claude/*
+!.claude/skills/
+!.claude/skills/**

 # Codex
 .codex/
@@ -114,17 +114,6 @@ Follow these rules for all code changes in this repository:
 - Keep comments and docstrings minimal and concise.
 - Assume the reader is familiar with vLLM.

-### Diagnosing CI failures
-
-Buildkite logs are public; no login needed. Details: [docs/contributing/ci/failures.md](docs/contributing/ci/failures.md).
-
-```bash
-# All failed-job logs for a PR's latest build (current branch's PR if omitted):
-.buildkite/scripts/ci-fetch-log.sh --pr <PR>
-# Any Buildkite build or job URL also works:
-.buildkite/scripts/ci-fetch-log.sh "<buildkite_url>"
-```
-
 ### Commit messages

 Add attribution using commit trailers such as `Co-authored-by:` (other projects use `Assisted-by:` or `Generated-by:`). For example:
@@ -319,82 +319,35 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
 endif()

 #
-# _C extension
+# Legacy _C extension (ROCm only — CUDA ops migrated to _C_stable_libtorch)
 #

-set(VLLM_EXT_SRC
-  "csrc/quantization/activation_kernels.cu"
-  "csrc/torch_bindings.cpp")
-
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
-
-  # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
-  set(CUTLASS_REVISION "v4.4.2")
-
-  # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
-  if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
-    set(VLLM_CUTLASS_SRC_DIR $ENV{VLLM_CUTLASS_SRC_DIR})
-  endif()
-
-  if(VLLM_CUTLASS_SRC_DIR)
-    if(NOT IS_ABSOLUTE VLLM_CUTLASS_SRC_DIR)
-      get_filename_component(VLLM_CUTLASS_SRC_DIR "${VLLM_CUTLASS_SRC_DIR}" ABSOLUTE)
-    endif()
-    message(STATUS "The VLLM_CUTLASS_SRC_DIR is set, using ${VLLM_CUTLASS_SRC_DIR} for compilation")
-    FetchContent_Declare(cutlass SOURCE_DIR ${VLLM_CUTLASS_SRC_DIR})
-  else()
-    FetchContent_Declare(
-        cutlass
-        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
-        # Please keep this in sync with CUTLASS_REVISION line above.
-        GIT_TAG ${CUTLASS_REVISION}
-        GIT_PROGRESS TRUE
-
-        # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
-        # Important: If GIT_SHALLOW is enabled then GIT_TAG works only with branch names and tags.
-        # So if the GIT_TAG above is updated to a commit hash, GIT_SHALLOW must be set to FALSE
-        GIT_SHALLOW TRUE
-    )
-  endif()
-  FetchContent_MakeAvailable(cutlass)
-
-  set_gencode_flags_for_srcs(
-    SRCS "${VLLM_EXT_SRC}"
-    CUDA_ARCHS "${CUDA_ARCHS}")
-
-# if CUDA endif
-endif()
-
-if (VLLM_GPU_LANG STREQUAL "HIP")
-  # Add QuickReduce kernels (ROCm-only; not part of stable ABI migration).
-  # TODO: Remove the cuda_view when ROCm upgrade to torch 2.11.
-  list(APPEND VLLM_EXT_SRC
+if(VLLM_GPU_LANG STREQUAL "HIP")
+  set(VLLM_EXT_SRC
+    "csrc/torch_bindings.cpp"
    "csrc/custom_quickreduce.cu"
    "csrc/cuda_view.cu"
-    "csrc/libtorch_stable/cuda_utils_kernels.cu"
-  )
-# if ROCM endif
-endif()
+    "csrc/libtorch_stable/cuda_utils_kernels.cu")

-message(STATUS "Enabling C extension.")
-define_extension_target(
-  _C
-  DESTINATION vllm
-  LANGUAGE ${VLLM_GPU_LANG}
-  SOURCES ${VLLM_EXT_SRC}
-  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
-  ARCHITECTURES ${VLLM_GPU_ARCHES}
-  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
-  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
-  USE_SABI 3
-  WITH_SOABI)
+  message(STATUS "Enabling C extension.")
+  define_extension_target(
+    _C
+    DESTINATION vllm
+    LANGUAGE ${VLLM_GPU_LANG}
+    SOURCES ${VLLM_EXT_SRC}
+    COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+    ARCHITECTURES ${VLLM_GPU_ARCHES}
+    INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+    INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
+    USE_SABI 3
+    WITH_SOABI)

-# If CUTLASS is compiled on NVCC >= 12.5, it by default uses
-# cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the
-# driver API. This causes problems when linking with earlier versions of CUDA.
-# Setting this variable sidesteps the issue by calling the driver directly.
-target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
+  # If CUTLASS is compiled on NVCC >= 12.5, it by default uses
+  # cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the
+  # driver API. This causes problems when linking with earlier versions of CUDA.
+  # Setting this variable sidesteps the issue by calling the driver directly.
+  target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
+endif() # _C HIP endif

 if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
  #
@@ -403,6 +356,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
  set(VLLM_STABLE_EXT_SRC
    "csrc/libtorch_stable/torch_bindings.cpp"
    "csrc/libtorch_stable/activation_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"
    "csrc/libtorch_stable/quantization/w8a8/fp8/per_token_group_quant.cu"
@@ -429,6 +383,38 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
    "csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu")

  if(VLLM_GPU_LANG STREQUAL "CUDA")
+    SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
+
+    # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
+    set(CUTLASS_REVISION "v4.4.2")
+
+    # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
+    if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
+      set(VLLM_CUTLASS_SRC_DIR $ENV{VLLM_CUTLASS_SRC_DIR})
+    endif()
+
+    if(VLLM_CUTLASS_SRC_DIR)
+      if(NOT IS_ABSOLUTE VLLM_CUTLASS_SRC_DIR)
+        get_filename_component(VLLM_CUTLASS_SRC_DIR "${VLLM_CUTLASS_SRC_DIR}" ABSOLUTE)
+      endif()
+      message(STATUS "The VLLM_CUTLASS_SRC_DIR is set, using ${VLLM_CUTLASS_SRC_DIR} for compilation")
+      FetchContent_Declare(cutlass SOURCE_DIR ${VLLM_CUTLASS_SRC_DIR})
+    else()
+      FetchContent_Declare(
+          cutlass
+          GIT_REPOSITORY https://github.com/nvidia/cutlass.git
+          # Please keep this in sync with CUTLASS_REVISION line above.
+          GIT_TAG ${CUTLASS_REVISION}
+          GIT_PROGRESS TRUE
+
+          # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
+          # Important: If GIT_SHALLOW is enabled then GIT_TAG works only with branch names and tags.
+          # So if the GIT_TAG above is updated to a commit hash, GIT_SHALLOW must be set to FALSE
+          GIT_SHALLOW TRUE
+      )
+    endif()
+    FetchContent_MakeAvailable(cutlass)
+
    list(APPEND VLLM_STABLE_EXT_SRC
      "csrc/libtorch_stable/cuda_view.cu"
      "csrc/libtorch_stable/cuda_utils_kernels.cu"
@@ -929,7 +915,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
      SRCS "${FP4_SM120_SRCS}"
      CUDA_ARCHS "${FP4_SM120_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${FP4_SM120_SRCS}")
-    target_compile_definitions(_C PRIVATE ENABLE_NVFP4_SM120=1)
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM120=1")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM120=1")
    message(STATUS "Building SM12x NVFP4 for archs: ${FP4_SM120_ARCHS}")
@@ -962,7 +947,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
      SRCS "${FP4_SM100_SRCS}"
      CUDA_ARCHS "${FP4_SM100_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${FP4_SM100_SRCS}")
-    target_compile_definitions(_C PRIVATE ENABLE_NVFP4_SM100=1)
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM100=1")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
    message(STATUS "Building SM10x/11x NVFP4/MXFP4 for archs: ${FP4_SM100_ARCHS}")
@@ -17,7 +17,7 @@ else()
  FetchContent_Declare(
    fmha_sm100
    GIT_REPOSITORY https://github.com/vllm-project/MSA.git
-    GIT_TAG 544eee5e09ae2dfa774d5b06739013f9b7402c57
+    GIT_TAG fee783153f3efe57e3e933c5cb7e267a7cebcfb5
    GIT_PROGRESS TRUE
    CONFIGURE_COMMAND ""
    BUILD_COMMAND ""
@@ -36,13 +36,38 @@ set(FMHA_SM100_PY_ROOT "${fmha_sm100_SOURCE_DIR}/python/fmha_sm100")

 install(FILES
  "${FMHA_SM100_PY_ROOT}/__init__.py"
+  "${FMHA_SM100_PY_ROOT}/api.py"
+  "${FMHA_SM100_PY_ROOT}/bench_utils.py"
+  "${FMHA_SM100_PY_ROOT}/jit.py"
  "${FMHA_SM100_PY_ROOT}/sparse.py"
+  "${FMHA_SM100_PY_ROOT}/sparse_fmha_adapter.py"
  DESTINATION vllm/third_party/fmha_sm100
  COMPONENT fmha_sm100)

+install(DIRECTORY "${FMHA_SM100_PY_ROOT}/csrc/"
+  DESTINATION vllm/third_party/fmha_sm100/csrc
+  COMPONENT fmha_sm100
+  PATTERN "__pycache__" EXCLUDE
+  PATTERN "*.pyc" EXCLUDE
+  PATTERN ".git*" EXCLUDE)
+
 install(DIRECTORY "${FMHA_SM100_PY_ROOT}/cute/"
  DESTINATION vllm/third_party/fmha_sm100/cute
  COMPONENT fmha_sm100
  PATTERN "__pycache__" EXCLUDE
  PATTERN "*.pyc" EXCLUDE
  PATTERN ".git*" EXCLUDE)
+
+install(DIRECTORY "${FMHA_SM100_PY_ROOT}/cutlass/include/"
+  DESTINATION vllm/third_party/fmha_sm100/cutlass/include
+  COMPONENT fmha_sm100
+  PATTERN "__pycache__" EXCLUDE
+  PATTERN "*.pyc" EXCLUDE
+  PATTERN ".git*" EXCLUDE)
+
+install(DIRECTORY "${FMHA_SM100_PY_ROOT}/cutlass/tools/util/include/"
+  DESTINATION vllm/third_party/fmha_sm100/cutlass/tools/util/include
+  COMPONENT fmha_sm100
+  PATTERN "__pycache__" EXCLUDE
+  PATTERN "*.pyc" EXCLUDE
+  PATTERN ".git*" EXCLUDE)
@@ -60,6 +60,7 @@ endif()

 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
  set(QUTLASS_SOURCES
+    csrc/qutlass_registration.cpp
    ${qutlass_SOURCE_DIR}/qutlass/csrc/bindings.cpp
    ${qutlass_SOURCE_DIR}/qutlass/csrc/gemm.cu
    ${qutlass_SOURCE_DIR}/qutlass/csrc/gemm_ada.cu
@@ -78,8 +79,19 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)

  if(CUTLASS_INCLUDE_DIR AND EXISTS "${CUTLASS_INCLUDE_DIR}/cutlass/cutlass.h")
    list(APPEND QUTLASS_INCLUDES "${CUTLASS_INCLUDE_DIR}")
+    if(CUTLASS_TOOLS_UTIL_INCLUDE_DIR AND
+       EXISTS "${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}/cutlass/util/packed_stride.hpp")
+      list(APPEND QUTLASS_INCLUDES "${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}")
+    else()
+      get_filename_component(_qutlass_cutlass_root "${CUTLASS_INCLUDE_DIR}" DIRECTORY)
+      if(EXISTS "${_qutlass_cutlass_root}/tools/util/include/cutlass/util/packed_stride.hpp")
+        list(APPEND QUTLASS_INCLUDES "${_qutlass_cutlass_root}/tools/util/include")
+      endif()
+    endif()
  elseif(EXISTS "${qutlass_SOURCE_DIR}/qutlass/third_party/cutlass/include/cutlass/cutlass.h")
-    list(APPEND QUTLASS_INCLUDES "${qutlass_SOURCE_DIR}/qutlass/third_party/cutlass/include")
+    list(APPEND QUTLASS_INCLUDES
+      "${qutlass_SOURCE_DIR}/qutlass/third_party/cutlass/include"
+      "${qutlass_SOURCE_DIR}/qutlass/third_party/cutlass/tools/util/include")
    message(STATUS "[QUTLASS] Using QuTLASS vendored CUTLASS headers (no vLLM CUTLASS detected).")
  else()
    message(FATAL_ERROR "[QUTLASS] CUTLASS headers not found. "
@@ -91,12 +103,23 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
    CUDA_ARCHS "${QUTLASS_ARCHS}"
  )

-  target_sources(_C PRIVATE ${QUTLASS_SOURCES})
-  target_include_directories(_C PRIVATE ${QUTLASS_INCLUDES})
-  target_compile_definitions(_C PRIVATE
+  # 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
+    LANGUAGE ${VLLM_GPU_LANG}
+    SOURCES ${QUTLASS_SOURCES}
+    COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+    ARCHITECTURES ${VLLM_GPU_ARCHES}
+    INCLUDE_DIRECTORIES ${QUTLASS_INCLUDES}
+    USE_SABI 3
+    WITH_SOABI)
+
+  target_compile_definitions(_qutlass_C PRIVATE
    QUTLASS_DISABLE_PYBIND=1
    TARGET_CUDA_ARCH=${QUTLASS_TARGET_CC}
-  )
+    CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)

  set_property(SOURCE ${QUTLASS_SOURCES} APPEND PROPERTY COMPILE_OPTIONS
    $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr --use_fast_math -O3>
@@ -111,4 +134,5 @@ else()
      "[QUTLASS] Skipping build: no supported arch (12.0f / 10.0f) found in "
      "CUDA_ARCHS='${CUDA_ARCHS}'.")
  endif()
+  add_custom_target(_qutlass_C)
 endif()
@@ -268,9 +268,14 @@ int64_t sm100_cutlass_mla_get_workspace_size(int64_t max_seq_len, int64_t num_ba
  using TileShapeD = typename MlaSm100Type::TileShapeD;
  arguments.problem_shape =
      cute::make_tuple(TileShapeH{}, static_cast<int>(max_seq_len), TileShapeD{}, static_cast<int>(num_batches));
-  // Assumes device 0 when getting sm_count.
-  arguments.hw_info.sm_count =
-      sm_count <= 0 ? cutlass::KernelHardwareInfo::query_device_multiprocessor_count(/*device_id=*/0) : sm_count;
+  if (sm_count <= 0) {
+    int current_device = 0;
+    cudaGetDevice(&current_device);
+    arguments.hw_info.sm_count =
+        cutlass::KernelHardwareInfo::query_device_multiprocessor_count(current_device);
+  } else {
+    arguments.hw_info.sm_count = sm_count;
+  }
  arguments.split_kv = static_cast<int>(num_kv_splits);
  MlaSm100Type::Fmha::set_split_kv(arguments);

@@ -67,6 +67,13 @@
  #include "../quantization/w8a8/fp8/nvidia/quant_utils.cuh"
 #endif

+// Direct float -> E4M3 FP8 conversion for the indexer Q / index-K outputs.
+#ifndef USE_ROCM
+  #include <cuda_fp8.h>
+#else
+  #include <hip/hip_fp8.h>
+#endif
+
 #ifndef FINAL_MASK
  #ifdef USE_ROCM
    #define FINAL_MASK 0xffffffffffffffffULL
@@ -75,6 +82,19 @@
  #endif
 #endif

+#ifdef USE_ROCM
+// ROCm-compatible direct float -> E4M3 FP8 conversion (mirrors the DeepSeek V4
+// fused kernel).
+__device__ __forceinline__ uint8_t rocm_cvt_float_to_fp8_e4m3(float val) {
+  #if defined(HIP_FP8_TYPE_OCP)
+  __hip_fp8_e4m3 fp8_val(val);
+  #else
+  __hip_fp8_e4m3_fnuz fp8_val(val);
+  #endif
+  return reinterpret_cast<uint8_t&>(fp8_val);
+}
+#endif
+
 namespace vllm {
 namespace minimax_m3_fused_ops {

@@ -193,6 +213,8 @@ __device__ __forceinline__ void storeElems(
  *reinterpret_cast<uint2*>(dst) = v;
 }

+// Main K/V cache store. kAuto = unquantized (cache_t == scalar_t); fp8 cache
+// dtypes use the scaled-convert path with identity scale.
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
 __device__ __forceinline__ void storeCacheElems(
    cache_t* __restrict__ dst, float const (&elems)[kElemsPerLane]) {
@@ -208,6 +230,32 @@ __device__ __forceinline__ void storeCacheElems(
  }
 }

+// Store 4 fp32 registers -> 4 contiguous E4M3 FP8 bytes (direct cast,
+// saturating to ±448). Used for the fp8 indexer-Q / index-K outputs; no scale
+// (RMSNorm outputs are O(1) and the score path only needs relative block
+// ordering).
+__device__ __forceinline__ void storeElemsFp8(
+    uint8_t* __restrict__ dst, float const (&elems)[kElemsPerLane]) {
+  constexpr float kFp8Max = 448.0f;
+#ifndef USE_ROCM
+  __nv_fp8x2_storage_t out2[kElemsPerLane / 2];
+  #pragma unroll
+  for (int i = 0; i < kElemsPerLane / 2; i++) {
+    float2 vv = make_float2(elems[2 * i], elems[2 * i + 1]);
+    vv.x = fminf(fmaxf(vv.x, -kFp8Max), kFp8Max);
+    vv.y = fminf(fmaxf(vv.y, -kFp8Max), kFp8Max);
+    out2[i] = __nv_cvt_float2_to_fp8x2(vv, __NV_SATFINITE, __NV_E4M3);
+  }
+  *reinterpret_cast<uint32_t*>(dst) = *reinterpret_cast<uint32_t const*>(out2);
+#else
+  #pragma unroll
+  for (int i = 0; i < kElemsPerLane; i++) {
+    float vv = fminf(fmaxf(elems[i], -kFp8Max), kFp8Max);
+    dst[i] = rocm_cvt_float_to_fp8_e4m3(vv);
+  }
+#endif
+}
+
 // ────────────────────────────────────────────────────────────────────────────
 // Kernel
 // ────────────────────────────────────────────────────────────────────────────
@@ -224,12 +272,14 @@ __device__ __forceinline__ void storeCacheElems(
 //     V : nkv  only if kInsertKV        (V-cache insert; no warps in dense)
 //     IQ: niq  only if kIsSparse        (norm+RoPE)
 //     IK: 1    only if kIsSparse        (norm+RoPE; +index-cache insert)
+// cache_t/kv_dt: main attention KV-cache dtype (auto/fp8). out_idx_t/kFp8Idx:
+// indexer index-K cache + index-Q output dtype (scalar_t or e4m3 byte).
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt,
-          bool kIsSparse, bool kInsertKV>
+          typename out_idx_t, bool kIsSparse, bool kInsertKV, bool kFp8Idx>
 __global__ void fusedMiniMaxM3QNormRopeKVInsertKernel(
    scalar_t* __restrict__ qkv,  // [N, qkv_row] in/out (packs index if sparse)
-    scalar_t* __restrict__ q_out,        // [N, nq*128] contiguous, or nullptr
-    scalar_t* __restrict__ index_q_out,  // [N, niq*128] contiguous, or nullptr
+    scalar_t* __restrict__ q_out,         // [N, nq*128] contiguous, or nullptr
+    out_idx_t* __restrict__ index_q_out,  // [N, niq*128]; scalar_t or e4m3 byte
    scalar_t const* __restrict__ q_norm_w,
    scalar_t const* __restrict__ k_norm_w,
    scalar_t const* __restrict__ iq_norm_w,
@@ -238,8 +288,8 @@ __global__ void fusedMiniMaxM3QNormRopeKVInsertKernel(
    int64_t const* __restrict__ positions,       // [N] i64
    int64_t const* __restrict__ slot_mapping,    // main K/V slots or nullptr
    int64_t const* __restrict__ index_slot_mapping,  // index K slots/nullptr
-    cache_t* __restrict__ kv_cache,      // [nb,2,bs,nkv,128] or nullptr
-    scalar_t* __restrict__ index_cache,  // [nb*bs, 128] or nullptr
+    cache_t* __restrict__ kv_cache,       // [nb,2,bs,nkv,128] or nullptr
+    out_idx_t* __restrict__ index_cache,  // [nb*bs, 128]; scalar_t or e4m3 byte
    float const eps, int const rotary_dim, int const num_tokens, int const nq,
    int const nkv, int const niq, int const block_size,
    // kv_cache strides (in elements) for logical shape [nb, 2, bs, nkv, 128].
@@ -334,9 +384,12 @@ __global__ void fusedMiniMaxM3QNormRopeKVInsertKernel(
      store_ptr = q_out + static_cast<int64_t>(tokenIdx) * nq * kHeadDim +
                  slot * kHeadDim;
    } else if (isIQ && index_q_out != nullptr) {
-      store_ptr = index_q_out +
-                  static_cast<int64_t>(tokenIdx) * niq * kHeadDim +
-                  (slot - iq_begin) * kHeadDim;
+      // bf16 index_q_out: gather here. fp8: written by the explicit fp8 store.
+      if constexpr (!kFp8Idx) {
+        store_ptr = index_q_out +
+                    static_cast<int64_t>(tokenIdx) * niq * kHeadDim +
+                    (slot - iq_begin) * kHeadDim;
+      }
    }

    // PDL: wait for the predecessor kernel (the qkv-projection GEMM that
@@ -356,7 +409,19 @@ __global__ void fusedMiniMaxM3QNormRopeKVInsertKernel(
      scalar_t const* cos_ptr = cos_sin_cache + pos * rotary_dim;
      normAndRope<scalar_t>(elems, laneId, eps, norm_w, do_rope, rotary_dim,
                            cos_ptr, /*apply_norm=*/norm_w != nullptr);
-      storeElems<scalar_t>(store_ptr + dim_base, elems);
+      if constexpr (kFp8Idx) {
+        // index_q is e4m3 bytes; Q/K (and in-place index_k) stay scalar_t.
+        if (isIQ && index_q_out != nullptr) {
+          storeElemsFp8(index_q_out +
+                            static_cast<int64_t>(tokenIdx) * niq * kHeadDim +
+                            (slot - iq_begin) * kHeadDim + dim_base,
+                        elems);
+        } else {
+          storeElems<scalar_t>(store_ptr + dim_base, elems);
+        }
+      } else {
+        storeElems<scalar_t>(store_ptr + dim_base, elems);
+      }
    }

    // ── Cache inserts (sparse serving only). ───────────────────────────────
@@ -367,8 +432,11 @@ __global__ void fusedMiniMaxM3QNormRopeKVInsertKernel(
                             : (isIK ? index_slot_mapping[tokenIdx] : -1);
      if (sm >= 0) {  // skip padded / unscheduled tokens
        if (isIK) {
-          scalar_t* dst = index_cache + sm * kHeadDim + dim_base;
-          storeElems<scalar_t>(dst, elems);
+          if constexpr (kFp8Idx) {
+            storeElemsFp8(index_cache + sm * kHeadDim + dim_base, elems);
+          } else {
+            storeElems<scalar_t>(index_cache + sm * kHeadDim + dim_base, elems);
+          }
        } else if (isK || isV) {
          // kv_cache logical shape [num_blocks, 2, block_size, nkv, head_dim].
          // Paging is logical (block = sm/block_size, token = sm%block_size);
@@ -398,19 +466,19 @@ __global__ void fusedMiniMaxM3QNormRopeKVInsertKernel(
 // Launch wrapper
 // ────────────────────────────────────────────────────────────────────────────
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
-void launchFusedMiniMaxM3(scalar_t* qkv, scalar_t* q_out, scalar_t* index_q_out,
-                          scalar_t const* q_norm_w, scalar_t const* k_norm_w,
-                          scalar_t const* iq_norm_w, scalar_t const* ik_norm_w,
-                          scalar_t const* cos_sin_cache,
-                          int64_t const* positions, int64_t const* slot_mapping,
-                          int64_t const* index_slot_mapping, cache_t* kv_cache,
-                          scalar_t* index_cache, float const eps,
-                          int const rotary_dim, int const num_tokens,
-                          int const nq, int const nkv, int const niq,
-                          int const block_size, int64_t const kv_s_block,
-                          int64_t const kv_s_kv, int64_t const kv_s_token,
-                          int64_t const kv_s_head, bool const has_index,
-                          bool const insert_kv, cudaStream_t stream) {
+void launchFusedMiniMaxM3(
+    scalar_t* qkv, scalar_t* q_out, void* index_q_out, scalar_t const* q_norm_w,
+    scalar_t const* k_norm_w, scalar_t const* iq_norm_w,
+    scalar_t const* ik_norm_w, scalar_t const* cos_sin_cache,
+    int64_t const* positions, int64_t const* slot_mapping,
+    int64_t const* index_slot_mapping, cache_t* kv_cache, void* index_cache,
+    float const eps, int const rotary_dim, int const num_tokens, int const nq,
+    int const nkv, int const niq, int const block_size,
+    int64_t const kv_s_block, int64_t const kv_s_kv, int64_t const kv_s_token,
+    int64_t const kv_s_head, bool const has_index, bool const insert_kv,
+    bool const fp8_idx, cudaStream_t stream) {
+  // Index outputs are scalar_t (bf16) or e4m3 bytes (uint8_t); reinterpret the
+  // void* pointers per instantiation in the LAUNCH macro.
  // Slot count must match the kernel's compile-time gating.
  int const v_slots = insert_kv ? nkv : 0;
  int const idx_slots = has_index ? niq + 1 : 0;
@@ -440,25 +508,27 @@ void launchFusedMiniMaxM3(scalar_t* qkv, scalar_t* q_out, scalar_t* index_q_out,
  config.attrs = attrs;
  config.numAttrs = (sm_version >= 90) ? 1 : 0;

-  #define LAUNCH(IS_SPARSE, INSERT)                                           \
-    cudaLaunchKernelEx(                                                       \
-        &config,                                                              \
-        fusedMiniMaxM3QNormRopeKVInsertKernel<scalar_t, cache_t, kv_dt,       \
-                                              IS_SPARSE, INSERT>,             \
-        qkv, q_out, index_q_out, q_norm_w, k_norm_w, iq_norm_w, ik_norm_w,    \
-        cos_sin_cache, positions, slot_mapping, index_slot_mapping, kv_cache, \
-        index_cache, eps, rotary_dim, num_tokens, nq, nkv, niq, block_size,   \
-        kv_s_block, kv_s_kv, kv_s_token, kv_s_head)
+  #define LAUNCH(IS_SPARSE, INSERT, FP8, OUT_T)                                \
+    cudaLaunchKernelEx(                                                        \
+        &config,                                                               \
+        fusedMiniMaxM3QNormRopeKVInsertKernel<scalar_t, cache_t, kv_dt, OUT_T, \
+                                              IS_SPARSE, INSERT, FP8>,         \
+        qkv, q_out, reinterpret_cast<OUT_T*>(index_q_out), q_norm_w, k_norm_w, \
+        iq_norm_w, ik_norm_w, cos_sin_cache, positions, slot_mapping,          \
+        index_slot_mapping, kv_cache, reinterpret_cast<OUT_T*>(index_cache),   \
+        eps, rotary_dim, num_tokens, nq, nkv, niq, block_size, kv_s_block,     \
+        kv_s_kv, kv_s_token, kv_s_head)
 #else
  // ROCm: standard kernel launch syntax (no PDL/stream serialization).
  // clang-format off
-  #define LAUNCH(IS_SPARSE, INSERT)                                          \
-    fusedMiniMaxM3QNormRopeKVInsertKernel<scalar_t, cache_t, kv_dt,          \
-                                          IS_SPARSE, INSERT>                 \
+  #define LAUNCH(IS_SPARSE, INSERT, FP8, OUT_T)                              \
+    fusedMiniMaxM3QNormRopeKVInsertKernel<scalar_t, cache_t, kv_dt, OUT_T,   \
+                                          IS_SPARSE, INSERT, FP8>            \
        <<<grid, kBlockSize, 0, stream>>>(                                   \
-            qkv, q_out, index_q_out, q_norm_w, k_norm_w, iq_norm_w,          \
-            ik_norm_w, cos_sin_cache, positions, slot_mapping,               \
-            index_slot_mapping, kv_cache, index_cache, eps, rotary_dim,      \
+            qkv, q_out, reinterpret_cast<OUT_T*>(index_q_out), q_norm_w,     \
+            k_norm_w, iq_norm_w, ik_norm_w, cos_sin_cache, positions,        \
+            slot_mapping, index_slot_mapping, kv_cache,                      \
+            reinterpret_cast<OUT_T*>(index_cache), eps, rotary_dim,          \
            num_tokens, nq, nkv, niq, block_size, kv_s_block, kv_s_kv,       \
            kv_s_token, kv_s_head)
  // clang-format on
@@ -466,14 +536,22 @@ void launchFusedMiniMaxM3(scalar_t* qkv, scalar_t* q_out, scalar_t* index_q_out,

  if (has_index) {
    if (insert_kv) {
-      LAUNCH(true, true);  // sparse serving
+      if (fp8_idx) {
+        LAUNCH(true, true, true, uint8_t);  // sparse serving, fp8 index outputs
+      } else {
+        LAUNCH(true, true, false, scalar_t);  // sparse serving, bf16
+      }
    } else {
-      LAUNCH(true, false);  // sparse profiling
+      if (fp8_idx) {
+        LAUNCH(true, false, true, uint8_t);  // sparse profiling, fp8 index_q
+      } else {
+        LAUNCH(true, false, false, scalar_t);  // sparse profiling, bf16
+      }
    }
  } else {
    // Dense layer: never has an index branch and never inserts here (the
    // generic Attention layer owns the KV insert).
-    LAUNCH(false, false);
+    LAUNCH(false, false, false, scalar_t);
  }
 #undef LAUNCH
 }
@@ -485,8 +563,9 @@ void launchFusedMiniMaxM3(scalar_t* qkv, scalar_t* q_out, scalar_t* index_q_out,
  vllm::minimax_m3_fused_ops::launchFusedMiniMaxM3<st, CACHE_T, KV_DTYPE>(     \
      reinterpret_cast<st*>(qkv.data_ptr()),                                   \
      q_out.has_value() ? reinterpret_cast<st*>(q_out->data_ptr()) : nullptr,  \
-      index_q_out.has_value() ? reinterpret_cast<st*>(index_q_out->data_ptr()) \
-                              : nullptr,                                       \
+      index_q_out.has_value()                                                  \
+          ? reinterpret_cast<void*>(index_q_out->data_ptr())                   \
+          : nullptr,                                                           \
      reinterpret_cast<st const*>(q_norm_weight.data_ptr()),                   \
      reinterpret_cast<st const*>(k_norm_weight.data_ptr()),                   \
      has_index ? reinterpret_cast<st const*>(index_q_norm_weight->data_ptr()) \
@@ -502,11 +581,11 @@ void launchFusedMiniMaxM3(scalar_t* qkv, scalar_t* q_out, scalar_t* index_q_out,
                : nullptr,                                                     \
      insert_kv ? reinterpret_cast<CACHE_T*>(kv_cache->data_ptr()) : nullptr,  \
      (insert_kv && has_index)                                                 \
-          ? reinterpret_cast<st*>(index_cache->data_ptr())                     \
+          ? reinterpret_cast<void*>(index_cache->data_ptr())                   \
          : nullptr,                                                           \
      static_cast<float>(eps), static_cast<int>(rotary_dim), num_tokens, nq,   \
      nkv, niq, static_cast<int>(block_size), kv_s_block, kv_s_kv, kv_s_token, \
-      kv_s_head, has_index, insert_kv, stream)
+      kv_s_head, has_index, insert_kv, fp8_idx, stream)

 // ────────────────────────────────────────────────────────────────────────────
 // Torch op wrapper
@@ -612,6 +691,7 @@ void fused_minimax_m3_qknorm_rope_kv_insert(
                 torch::headeronly::ScalarType::Long &&
             index_slot_mapping->numel() == slot_mapping->numel()),
        "index_slot_mapping must be int64 CUDA with slot_mapping length");
+    // Main attention KV cache: auto matches qkv, fp8 uses uint8 storage.
    if (kv_dt == vllm::Fp8KVCacheDataType::kAuto) {
      STD_TORCH_CHECK(kv_cache->scalar_type() == qkv.scalar_type(),
                      "auto kv_cache dtype must match qkv");
@@ -620,9 +700,13 @@ void fused_minimax_m3_qknorm_rope_kv_insert(
          kv_cache->scalar_type() == torch::headeronly::ScalarType::Byte,
          "fp8 kv_cache must use uint8 storage");
    }
-    STD_TORCH_CHECK(index_cache.has_value() &&
-                        index_cache->scalar_type() == qkv.scalar_type(),
-                    "insert mode requires matching index_cache");
+    // Indexer index-K cache: independent dtype -- qkv dtype or fp8 e4m3.
+    STD_TORCH_CHECK(
+        index_cache.has_value() &&
+            (index_cache->scalar_type() == qkv.scalar_type() ||
+             index_cache->scalar_type() ==
+                 torch::headeronly::ScalarType::Float8_e4m3fn),
+        "insert mode requires index_cache matching qkv dtype or fp8 e4m3");
    STD_TORCH_CHECK(kv_cache->dim() == 5 && kv_cache->stride(4) == 1,
                    "kv_cache must be [nb,2,bs,nkv,head_dim] with contiguous "
                    "head_dim (stride(4)==1)");
@@ -652,14 +736,31 @@ void fused_minimax_m3_qknorm_rope_kv_insert(
        "index_q_out requires the index branch (num_index_heads > 0)");
    STD_TORCH_CHECK(
        index_q_out->is_cuda() && index_q_out->is_contiguous() &&
-            index_q_out->scalar_type() == qkv.scalar_type(),
-        "index_q_out must be a contiguous CUDA tensor matching qkv dtype");
+            (index_q_out->scalar_type() == qkv.scalar_type() ||
+             index_q_out->scalar_type() ==
+                 torch::headeronly::ScalarType::Float8_e4m3fn),
+        "index_q_out must be contiguous CUDA, qkv dtype or fp8 e4m3");
    STD_TORCH_CHECK(index_q_out->numel() ==
                        static_cast<int64_t>(num_tokens) * niq * kHeadDim,
                    "index_q_out must have num_tokens * num_index_heads * 128 "
                    "elements");
  }

+  // fp8 index path: the index-K cache and index-Q outputs are e4m3 bytes while
+  // q/k/v + q_out stay qkv dtype. Both index outputs must agree.
+  auto const kFp8 = torch::headeronly::ScalarType::Float8_e4m3fn;
+  bool const fp8_idx =
+      (index_cache.has_value() && index_cache->scalar_type() == kFp8) ||
+      (index_q_out.has_value() && index_q_out->scalar_type() == kFp8);
+  if (fp8_idx) {
+    STD_TORCH_CHECK(
+        !index_cache.has_value() || index_cache->scalar_type() == kFp8,
+        "fp8 index path: index_cache must be fp8 e4m3");
+    STD_TORCH_CHECK(
+        !index_q_out.has_value() || index_q_out->scalar_type() == kFp8,
+        "fp8 index path: index_q_out must be fp8 e4m3");
+  }
+
  const torch::stable::accelerator::DeviceGuard device_guard(
      qkv.get_device_index());
  auto stream = get_current_cuda_stream(qkv.get_device_index());
@@ -81,11 +81,11 @@ __global__ void rms_norm_kernel(
 #pragma unroll
    for (int j = 0; j < VEC_SIZE; j++) {
      float x = static_cast<float>(src1.val[j]);
-      scalar_t normalized = static_cast<scalar_t>(x * s_variance);
      if constexpr (HasWeight) {
-        dst.val[j] = normalized * src2.val[j];
+        float w = static_cast<float>(src2.val[j]);
+        dst.val[j] = static_cast<scalar_t>(x * s_variance * w);
      } else {
-        dst.val[j] = normalized;
+        dst.val[j] = static_cast<scalar_t>(x * s_variance);
      }
    }
    v_out[i] = dst;
@@ -151,7 +151,8 @@ fused_add_rms_norm_kernel(
 #pragma unroll
      for (int j = 0; j < width; ++j) {
        float x = Converter::convert(res.data[j]);
-        out.data[j] = Converter::convert(x * s_variance) * w.data[j];
+        float wf = Converter::convert(w.data[j]);
+        out.data[j] = Converter::convert(x * s_variance * wf);
      }
    } else {
 #pragma unroll
@@ -198,8 +199,8 @@ fused_add_rms_norm_kernel(
  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
    float x = (float)residual[blockIdx.x * hidden_size + idx];
    if constexpr (HasWeight) {
-      input[blockIdx.x * input_stride + idx] =
-          (scalar_t)(x * s_variance) * weight[idx];
+      float w = (float)weight[idx];
+      input[blockIdx.x * input_stride + idx] = (scalar_t)(x * s_variance * w);
    } else {
      input[blockIdx.x * input_stride + idx] = (scalar_t)(x * s_variance);
    }
@@ -66,8 +66,13 @@ __global__ void rms_norm_static_fp8_quant_kernel(
 #pragma unroll
    for (int j = 0; j < VEC_SIZE; j++) {
      float x = static_cast<float>(src1.val[j]);
-      // Multiply in weight's native dtype to match rms_norm_kernel.
-      scalar_t out_norm = static_cast<scalar_t>(x * s_variance) * src2.val[j];
+      float w = static_cast<float>(src2.val[j]);
+      // Round normalized result through scalar_t to match the precision of the
+      // unfused composite (rms_norm writes scalar_t, then
+      // static_scaled_fp8_quant re-loads it as float before FP8 conversion).
+      // Without this round, the fused path is strictly more accurate and
+      // disagrees with the composite at exact E4M3 quantization tie boundaries.
+      scalar_t out_norm = static_cast<scalar_t>(x * s_variance * w);
      out[blockIdx.x * hidden_size + idx * VEC_SIZE + j] =
          scaled_fp8_conversion<true, fp8_type>(static_cast<float>(out_norm),
                                                scale_inv);
@@ -137,8 +142,12 @@ fused_add_rms_norm_static_fp8_quant_kernel(
 #pragma unroll
    for (int i = 0; i < width; ++i) {
      float x = Converter::convert(res.data[i]);
-      // Multiply in weight's native dtype to match fused_add_rms_norm_kernel.
-      HipT out_norm_h = Converter::convert(x * s_variance) * w.data[i];
+      float wf = Converter::convert(w.data[i]);
+      // See note in rms_norm_static_fp8_quant_kernel: round through scalar_t
+      // to match the unfused composite path at FP8 boundaries. We use the
+      // backend's hip_type for the intermediate since c10::Half/BFloat16 has
+      // ambiguous conversions on CUDA and no implicit conversion on ROCm.
+      HipT out_norm_h = Converter::convert(x * s_variance * wf);
      out[id * width + i] = scaled_fp8_conversion<true, fp8_type>(
          Converter::convert(out_norm_h), scale_inv);
    }
@@ -183,8 +192,10 @@ fused_add_rms_norm_static_fp8_quant_kernel(

  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
    float x = (float)residual[blockIdx.x * hidden_size + idx];
-    // Multiply in weight's native dtype to match fused_add_rms_norm_kernel.
-    scalar_t out_norm = static_cast<scalar_t>(x * s_variance) * weight[idx];
+    float w = (float)weight[idx];
+    // See note in rms_norm_static_fp8_quant_kernel: round through scalar_t
+    // to match the unfused composite path at FP8 boundaries.
+    scalar_t out_norm = static_cast<scalar_t>(x * s_variance * w);
    out[blockIdx.x * hidden_size + idx] = scaled_fp8_conversion<true, fp8_type>(
        static_cast<float>(out_norm), scale_inv);
  }
@@ -9,7 +9,7 @@
 #include <torch/headeronly/core/ScalarType.h>

 #include "../../cuda_compat.h"
-#include "core/math.hpp"
+#include "libtorch_stable/core/math.hpp"
 #include "libtorch_stable/dispatch_utils.h"
 #include "libtorch_stable/torch_utils.h"

@@ -2,9 +2,25 @@

 #include <torch/csrc/stable/library.h>
 #include <torch/csrc/stable/tensor.h>
+#include <torch/headeronly/util/Exception.h>

 #include <optional>
 #include <string>
+#include <vector>
+
+#include <torch/csrc/stable/ops.h>
+
+inline torch::stable::Tensor weak_ref_tensor(torch::stable::Tensor& tensor) {
+  // Ensure tensor is on CUDA
+  STD_TORCH_CHECK(tensor.device().is_cuda(), "Tensor must be on CUDA device");
+
+  // Get the raw data pointer
+  void* data_ptr = tensor.mutable_data_ptr();
+
+  /// Create a new tensor from the raw data pointer
+  return torch::stable::from_blob(data_ptr, tensor.sizes(), tensor.strides(),
+                                  tensor.device(), tensor.scalar_type());
+}

 void per_token_group_quant_fp8(const torch::stable::Tensor& input,
                               torch::stable::Tensor& output_q,
@@ -371,6 +387,18 @@ void silu_and_mul(torch::stable::Tensor& out, torch::stable::Tensor& input);
 void silu_and_mul_clamp(torch::stable::Tensor& out,
                        torch::stable::Tensor& input, double limit,
                        double alpha = 1.0, double beta = 0.0);
+
+void silu_and_mul_quant(torch::stable::Tensor& out,
+                        torch::stable::Tensor& input,
+                        torch::stable::Tensor& scale);
+
+void persistent_masked_m_silu_mul_quant(
+    const torch::stable::Tensor& input,              // (E, T, 2*H)
+    const torch::stable::Tensor& tokens_per_expert,  // (E)
+    torch::stable::Tensor& y_q,                      // (E, T, H) [OUT]
+    torch::stable::Tensor& y_s,  // (E, T, H//group_size) [OUT]
+    bool use_ue8m0);
+
 void mul_and_silu(torch::stable::Tensor& out, torch::stable::Tensor& input);
 void gelu_and_mul(torch::stable::Tensor& out, torch::stable::Tensor& input);
 void gelu_tanh_and_mul(torch::stable::Tensor& out,
@@ -1,16 +1,12 @@
-#include <ATen/cuda/CUDAContext.h>
-#include <torch/all.h>
-#include <c10/cuda/CUDAGuard.h>
+#include "libtorch_stable/torch_utils.h"

 #include <cmath>
-#include "core/math.hpp"
-#include "../cuda_compat.h"
-#include "dispatch_utils.h"

+#include "libtorch_stable/core/math.hpp"
+#include "cuda_compat.h"
+#include "libtorch_stable/dispatch_utils.h"
 #include "quantization/w8a8/fp8/common.cuh"

-#include <c10/util/Float8_e4m3fn.h>
-
 #ifndef USE_ROCM
  #include <cuda_bf16.h>
  #include <cuda_fp16.h>
@@ -33,7 +29,6 @@ typedef __hip_fp8x4_e4m3_fnuz __nv_fp8x4_e4m3;
  #endif
 #endif

-#include "core/registration.h"
 namespace vllm {

 template <typename T>
@@ -564,41 +559,47 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
 }  // namespace vllm

 // Launch activation, gating, and quantize kernel.
-#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL)                               \
-  int d = input.size(-1) / 2;                                               \
-  int64_t num_tokens = input.numel() / input.size(-1);                      \
-  dim3 grid(num_tokens, num_tokens > 16 ? num_tokens > 32 ? 1 : 2 : 4);     \
-  dim3 block(std::min(d, 512));                                             \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));         \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();             \
-  VLLM_DISPATCH_FLOATING_TYPES(                                             \
-      input.scalar_type(), "act_and_mul_kernel", [&] {                      \
-        VLLM_DISPATCH_FP8_TYPES(                                            \
-            out.scalar_type(), "fused_add_rms_norm_kernel_fp8_type", [&] {  \
-              vllm::act_and_mul_quant_kernel<scalar_t, KERNEL<scalar_t>,    \
-                                             fp8_t>                         \
-                  <<<grid, block, 0, stream>>>(out.data_ptr<fp8_t>(),       \
-                                               input.data_ptr<scalar_t>(),  \
-                                               scale.data_ptr<float>(), d); \
-            });                                                             \
+#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL)                             \
+  int d = input.size(-1) / 2;                                             \
+  int64_t num_tokens = input.numel() / input.size(-1);                    \
+  dim3 grid(num_tokens, num_tokens > 16 ? num_tokens > 32 ? 1 : 2 : 4);   \
+  dim3 block(std::min(d, 512));                                           \
+  const torch::stable::accelerator::DeviceGuard device_guard(             \
+      input.get_device_index());                                          \
+  const cudaStream_t stream =                                             \
+      get_current_cuda_stream(input.get_device_index());                  \
+  VLLM_STABLE_DISPATCH_FLOATING_TYPES(                                    \
+      input.scalar_type(), "act_and_mul_kernel", [&] {                    \
+        VLLM_STABLE_DISPATCH_FP8_TYPES(                                   \
+            out.scalar_type(), "act_and_mul_quant_kernel_fp8_type", [&] { \
+              vllm::act_and_mul_quant_kernel<scalar_t, KERNEL<scalar_t>,  \
+                                             fp8_t>                       \
+                  <<<grid, block, 0, stream>>>(                           \
+                      out.mutable_data_ptr<fp8_t>(),                      \
+                      input.const_data_ptr<scalar_t>(),                   \
+                      scale.const_data_ptr<float>(), d);                  \
+            });                                                           \
      });

-void silu_and_mul_quant(torch::Tensor& out,    // [..., d]
-                        torch::Tensor& input,  // [..., 2 * d]
-                        torch::Tensor& scale) {
-  TORCH_CHECK(out.dtype() == torch::kFloat8_e4m3fn ||
-              out.dtype() == torch::kFloat8_e4m3fnuz);
-  TORCH_CHECK(input.dtype() == torch::kFloat16 ||
-              input.dtype() == torch::kBFloat16);
-  TORCH_CHECK(input.size(-1) % 2 == 0);
+void silu_and_mul_quant(torch::stable::Tensor& out,    // [..., d]
+                        torch::stable::Tensor& input,  // [..., 2 * d]
+                        torch::stable::Tensor& scale) {
+  STD_TORCH_CHECK(
+      out.scalar_type() == torch::headeronly::ScalarType::Float8_e4m3fn ||
+      out.scalar_type() == torch::headeronly::ScalarType::Float8_e4m3fnuz);
+  STD_TORCH_CHECK(
+      input.scalar_type() == torch::headeronly::ScalarType::Half ||
+          input.scalar_type() == torch::headeronly::ScalarType::BFloat16,
+      "Input must be FP16 or BF16");
+  STD_TORCH_CHECK(input.size(-1) % 2 == 0);
  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel);
 }

 void persistent_masked_m_silu_mul_quant(
-    const at::Tensor& input,              // (E, T, 2*H)
-    const at::Tensor& tokens_per_expert,  // (E)
-    at::Tensor& y_q,                      // (E, T, H) [OUT]
-    at::Tensor& y_s,                      // (E, T, H//group_size) [OUT]
+    const torch::stable::Tensor& input,              // (E, T, 2*H)
+    const torch::stable::Tensor& tokens_per_expert,  // (E)
+    torch::stable::Tensor& y_q,                      // (E, T, H) [OUT]
+    torch::stable::Tensor& y_s,  // (E, T, H//group_size) [OUT]
    bool cast_scale_ue8m0) {
 #ifndef USE_ROCM

@@ -606,14 +607,18 @@ void persistent_masked_m_silu_mul_quant(
  // fixed GROUP_SIZE of 128.
  static constexpr int GROUP_SIZE = 128;

-  TORCH_CHECK(input.dtype() == torch::kBFloat16);
-  TORCH_CHECK(y_q.dtype() == torch::kFloat8_e4m3fn ||
-              y_q.dtype() == torch::kFloat8_e4m3fnuz);
-  TORCH_CHECK(input.size(-1) % (GROUP_SIZE * 2) == 0);
+  STD_TORCH_CHECK(input.scalar_type() ==
+                  torch::headeronly::ScalarType::BFloat16);
+  STD_TORCH_CHECK(
+      y_q.scalar_type() == torch::headeronly::ScalarType::Float8_e4m3fn ||
+      y_q.scalar_type() == torch::headeronly::ScalarType::Float8_e4m3fnuz);
+  STD_TORCH_CHECK(input.size(-1) % (GROUP_SIZE * 2) == 0);

  bool const is_packed_ue8m0 =
-      (y_s.dtype() == torch::kInt32 && cast_scale_ue8m0);
-  TORCH_CHECK(y_s.dtype() == torch::kFloat32 || is_packed_ue8m0);
+      (y_s.scalar_type() == torch::headeronly::ScalarType::Int &&
+       cast_scale_ue8m0);
+  STD_TORCH_CHECK(y_s.scalar_type() == torch::headeronly::ScalarType::Float ||
+                  is_packed_ue8m0);

  using Idx_t = int64_t;

@@ -631,7 +636,7 @@ void persistent_masked_m_silu_mul_quant(

  int const NUM_GROUPS = H / GROUP_SIZE;

-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  const cudaStream_t stream = get_current_cuda_stream(input.get_device_index());

  // TODO: Get this from cuda_arch ?
  static constexpr int SILU_V2_BLOCK_COUNT = 132 * 32;
@@ -643,18 +648,21 @@ void persistent_masked_m_silu_mul_quant(
    static constexpr int max_shared_mem_bytes =                                \
        GROUP_SIZE * 2 * STAGES * NUM_WARPS * 2;                               \
    dim3 grid(sms), block(THREAD_COUNT);                                       \
-    const at::cuda::OptionalCUDAGuard device_guard(device_of(input));          \
-    VLLM_DISPATCH_FP8_TYPES(                                                   \
+    const torch::stable::accelerator::DeviceGuard device_guard(                \
+        input.get_device_index());                                             \
+    VLLM_STABLE_DISPATCH_FP8_TYPES(                                            \
        y_q.scalar_type(), "silu_mul_fp8_quant_deep_gemm_kernel", [&] {        \
          vllm::silu_mul_fp8_quant_deep_gemm_kernel<                           \
              BLOCK_COUNT, max_shared_mem_bytes, fp8_t, scale_t, THREAD_COUNT, \
              Idx_t, CEIL_UE8M0, GROUP_SIZE, STAGES>                           \
              <<<grid, block, max_shared_mem_bytes + (E + 1) * 16, stream>>>(  \
-                  reinterpret_cast<__nv_bfloat16*>(input.data_ptr()),          \
-                  (fp8_t*)y_q.data_ptr(),                                      \
-                  reinterpret_cast<scale_t*>(y_s.data_ptr()),                  \
-                  reinterpret_cast<int32_t*>(tokens_per_expert.data_ptr()), E, \
-                  T, H, stride_i_e, stride_i_t, stride_i_h, stride_yq_e,       \
+                  reinterpret_cast<const __nv_bfloat16*>(                      \
+                      input.const_data_ptr()),                                 \
+                  y_q.mutable_data_ptr<fp8_t>(),                               \
+                  reinterpret_cast<scale_t*>(y_s.mutable_data_ptr()),          \
+                  reinterpret_cast<const int32_t*>(                            \
+                      tokens_per_expert.const_data_ptr()),                     \
+                  E, T, H, stride_i_e, stride_i_t, stride_i_h, stride_yq_e,    \
                  stride_yq_t, stride_yq_h, STRIDE_YS_E, STRIDE_YS_T,          \
                  STRIDE_YS_G, STRIDE_YS_P, stride_counts_e);                  \
        });
@@ -679,7 +687,7 @@ void persistent_masked_m_silu_mul_quant(
  Idx_t stride_ys_g = y_s.stride(2);
  Idx_t stride_ys_p = 0;
  if (!cast_scale_ue8m0) {
-    TORCH_CHECK(!is_packed_ue8m0);
+    STD_TORCH_CHECK(!is_packed_ue8m0);
    LAUNCH_ON_H(float, stride_ys_e, stride_ys_t, stride_ys_g, stride_ys_p,
                false);
    return;
@@ -692,8 +700,8 @@ void persistent_masked_m_silu_mul_quant(
    return;
  }

-  TORCH_CHECK(cast_scale_ue8m0 && is_packed_ue8m0);
-  TORCH_CHECK(y_s.dtype() == torch::kInt32);
+  STD_TORCH_CHECK(cast_scale_ue8m0 && is_packed_ue8m0);
+  STD_TORCH_CHECK(y_s.scalar_type() == torch::headeronly::ScalarType::Int);

  // Int32 packed ue8m0 scales tensor.
  // Let E, T, G be the number to experts, number of tokens and number of groups
@@ -31,7 +31,7 @@

 #include "cutlass/util/packed_stride.hpp"

-#include "core/math.hpp"
+#include "libtorch_stable/core/math.hpp"
 #include "core/batch_invariant.hpp"

 using namespace cute;
@@ -31,7 +31,7 @@

 #include "cutlass/util/packed_stride.hpp"

-#include "core/math.hpp"
+#include "libtorch_stable/core/math.hpp"
 #include "core/batch_invariant.hpp"

 using namespace cute;
@@ -19,7 +19,7 @@
 #include "cutlass/gemm/collective/collective_builder.hpp"
 #include "cutlass/util/packed_stride.hpp"

-#include "core/math.hpp"
+#include "libtorch_stable/core/math.hpp"
 #include "libtorch_stable/cutlass_extensions/common.hpp"
 // clang-format on

@@ -14,7 +14,7 @@
 #include "cutlass/epilogue/collective/collective_builder.hpp"
 #include "cutlass/gemm/collective/collective_builder.hpp"

-#include "core/math.hpp"
+#include "libtorch_stable/core/math.hpp"
 #include "libtorch_stable/cutlass_extensions/common.hpp"
 // clang-format on

@@ -22,7 +22,7 @@
 #include "cutlass/epilogue/threadblock/fusion/visitors.hpp"
 #include "cutlass/gemm/kernel/default_gemm_universal_with_visitor.h"

-#include "core/math.hpp"
+#include "libtorch_stable/core/math.hpp"
 #include "libtorch_stable/cutlass_extensions/common.hpp"
 // clang-format on

@@ -301,8 +301,9 @@ __global__ void per_token_group_quant_8bit_packed_register_kernel(

  const int sf_k_local = local_group_id % kGroupsPerBlockX;
  const int row_local = local_group_id / kGroupsPerBlockX;
-  const int sf_k_idx = blockIdx.x * kGroupsPerBlockX + sf_k_local;
-  const int mn_idx = blockIdx.y * kRowsPerBlock + row_local;
+  // Rows on grid.x: mn scales with tokens and can exceed the 65535 grid.y cap.
+  const int sf_k_idx = blockIdx.y * kGroupsPerBlockX + sf_k_local;
+  const int mn_idx = blockIdx.x * kRowsPerBlock + row_local;

 #if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
  asm volatile("griddepcontrol.wait;");
@@ -496,14 +497,15 @@ void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
                  " is not a multiple of 4.");
  const int kx = GetGroupsPerBlockX(padded_groups_per_row);
  const int ry = 16 / kx;
-  const int64_t blocks_x = padded_groups_per_row / kx;
-  const int64_t blocks_y = (tma_aligned_mn + ry - 1) / ry;
+  const int64_t row_blocks = (tma_aligned_mn + ry - 1) / ry;
+  const int64_t sf_k_blocks = padded_groups_per_row / kx;
  const int num_threads = (kx * ry) * THREADS_PER_GROUP;
-  // CUDA caps grid.x and grid.y at 2^31 - 1; guard against pathological inputs.
-  STD_TORCH_CHECK(blocks_x <= static_cast<int64_t>(INT32_MAX) &&
-                      blocks_y <= static_cast<int64_t>(INT32_MAX),
+  // CUDA caps grid.x at 2^31 - 1 and grid.y at 2^16 - 1 (65535).
+  constexpr int64_t kMaxGridDimYZ = 65535;
+  STD_TORCH_CHECK(row_blocks <= static_cast<int64_t>(INT32_MAX) &&
+                      sf_k_blocks <= kMaxGridDimYZ,
                  "per_token_group_quant_8bit_packed grid too large: (",
-                  blocks_x, ", ", blocks_y, ").");
+                  row_blocks, ", ", sf_k_blocks, ").");

  auto dst_type = output_q.scalar_type();

@@ -513,8 +515,8 @@ void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
  #define LAUNCH_REG_KERNEL_INST(T, DST_DTYPE, KX, RY)                         \
    do {                                                                       \
      cudaLaunchConfig_t config = {};                                          \
-      config.gridDim = dim3(static_cast<unsigned int>(blocks_x),               \
-                            static_cast<unsigned int>(blocks_y));              \
+      config.gridDim = dim3(static_cast<unsigned int>(row_blocks),             \
+                            static_cast<unsigned int>(sf_k_blocks));           \
      config.blockDim = dim3(num_threads);                                     \
      config.dynamicSmemBytes = 0;                                             \
      config.stream = stream;                                                  \
@@ -539,8 +541,8 @@ void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
 #else
  #define LAUNCH_REG_KERNEL_INST(T, DST_DTYPE, KX, RY)                         \
    do {                                                                       \
-      dim3 grid(static_cast<unsigned int>(blocks_x),                           \
-                static_cast<unsigned int>(blocks_y));                          \
+      dim3 grid(static_cast<unsigned int>(row_blocks),                         \
+                static_cast<unsigned int>(sf_k_blocks));                       \
      dim3 block(num_threads);                                                 \
      per_token_group_quant_8bit_packed_register_kernel<T, DST_DTYPE, 128, KX, \
                                                        RY>                    \
@@ -34,6 +34,20 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
  // TODO: Remove this once ROCm upgrade to torch 2.11.
  ops.def("get_cuda_view_from_cpu_tensor(Tensor cpu_tensor) -> Tensor");

+  // Note about marlin kernel 'workspace' arguments:
+  // Technically these should be mutable since they are modified by the kernel.
+  // But since they are set back to zero once the kernel is finished we can
+  // hand wave and say that they have no net effect.
+  //
+  // The reason to mark 'workspace' as immutable is so that they don't interfere
+  // with using ScalarType arguments in the ops. If they are marked as mutable,
+  // pytorch throws an assert in
+  // 'torch._higher_order_ops._register_effectful_op' that prevents these
+  // kernels from being torch.compile'd.
+  // See the following document for more info on custom types and ops that use
+  // custom types:
+  // https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA
+
  // Machete (Dense) Optimized Mixed Precision GEMM for Hopper.
  ops.def(
      "machete_supported_schedules("
@@ -480,6 +494,11 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
      "Tensor workspace, int k, int max_seq_len) -> ()");

  // Activation ops
+  ops.def(
+      "persistent_masked_m_silu_mul_quant(Tensor input, Tensor counts, Tensor! "
+      "y_q, Tensor! y_s, bool use_ue8m0) -> ()");
+  ops.def("weak_ref_tensor(Tensor input) -> Tensor");
+
  // Activation function used in SwiGLU.
  ops.def("silu_and_mul(Tensor! result, Tensor input) -> ()");

@@ -492,6 +511,10 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
      "silu_and_mul_with_clamp(Tensor! result, Tensor input, float limit, "
      "float alpha=1.0, float beta=0.0) -> ()");

+  // SwiGLU activation with FP8 quantization.
+  ops.def(
+      "silu_and_mul_quant(Tensor! result, Tensor input, Tensor scale) -> ()");
+
  // Activation function used in GeGLU with `none` approximation.
  ops.def("gelu_and_mul(Tensor! out, Tensor input) -> ()");

@@ -690,6 +713,10 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
  ops.impl("persistent_topk", TORCH_BOX(&persistent_topk));

  // Activation kernels (shared CUDA/ROCm)
+  ops.impl("persistent_masked_m_silu_mul_quant",
+           TORCH_BOX(&persistent_masked_m_silu_mul_quant));
+  ops.impl("weak_ref_tensor", TORCH_BOX(&weak_ref_tensor));
+  ops.impl("silu_and_mul_quant", TORCH_BOX(&silu_and_mul_quant));
  ops.impl("silu_and_mul", TORCH_BOX(&silu_and_mul));
  ops.impl("mul_and_silu", TORCH_BOX(&mul_and_silu));
  ops.impl("gelu_and_mul", TORCH_BOX(&gelu_and_mul));
@@ -9,28 +9,6 @@

 #include <vector>

-torch::Tensor weak_ref_tensor(torch::Tensor& tensor) {
-  // Ensure tensor is on CUDA
-  if (!tensor.is_cuda()) {
-    throw std::runtime_error("Tensor must be on CUDA device");
-  }
-
-  // Get the raw data pointer
-  void* data_ptr = tensor.data_ptr();
-
-  // Get tensor sizes and strides
-  std::vector<int64_t> sizes = tensor.sizes().vec();
-  std::vector<int64_t> strides = tensor.strides().vec();
-
-  // Get tensor options (dtype, device)
-  auto options = tensor.options();
-
-  // Create a new tensor from the raw data pointer
-  auto new_tensor = torch::from_blob(data_ptr, sizes, strides, options);
-
-  return new_tensor;
-}
-
 // rms_norm and fused_add_rms_norm declarations also exist in
 // csrc/libtorch_stable/ops.h (torch::stable ABI for CUDA). They remain here
 // because the CPU build still uses these torch::Tensor declarations.
@@ -53,16 +31,6 @@ void silu_and_mul(torch::Tensor& out, torch::Tensor& input);
 void silu_and_mul_clamp(torch::Tensor& out, torch::Tensor& input, double limit,
                        double alpha = 1.0, double beta = 0.0);

-void silu_and_mul_quant(torch::Tensor& out, torch::Tensor& input,
-                        torch::Tensor& scale);
-
-void persistent_masked_m_silu_mul_quant(
-    const at::Tensor& input,   // (E, T, 2*H)
-    const at::Tensor& counts,  // (E)
-    at::Tensor& y_q,           // (E, T, H) [OUT]
-    at::Tensor& y_s,           // (E, T, H//group_size) [OUT]
-    bool use_ue8m0);
-
 void gelu_and_mul(torch::Tensor& out, torch::Tensor& input);

 void gelu_tanh_and_mul(torch::Tensor& out, torch::Tensor& input);
@@ -0,0 +1,5 @@
+#include "core/registration.h"
+
+// QuTLASS registers torch.ops._qutlass_C via TORCH_LIBRARY in bindings.cpp.
+// This stub lets Python import vllm._qutlass_C to trigger op registration.
+REGISTER_EXTENSION(_qutlass_C)
@@ -20,17 +20,6 @@

 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  // vLLM custom ops
-  //
-
-  ops.def(
-      "persistent_masked_m_silu_mul_quant(Tensor input, Tensor counts, Tensor! "
-      "y_q, Tensor! y_s,"
-      "bool use_ue8m0) -> ()");
-  ops.impl("persistent_masked_m_silu_mul_quant", torch::kCUDA,
-           &persistent_masked_m_silu_mul_quant);
-
-  ops.def("weak_ref_tensor(Tensor input) -> Tensor");
-  ops.impl("weak_ref_tensor", torch::kCUDA, &weak_ref_tensor);

 #ifdef USE_ROCM
  // TODO: Remove this once we upgrade to torch 2.11.
@@ -39,35 +28,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def("get_cuda_view_from_cpu_tensor(Tensor cpu_tensor) -> Tensor");
  ops.impl("get_cuda_view_from_cpu_tensor", torch::kCPU,
           &get_cuda_view_from_cpu_tensor);
-#endif
-
-  // Activation ops (quantized only — basic ops moved to _C_stable_libtorch)
-  ops.def(
-      "silu_and_mul_quant(Tensor! result, Tensor input, Tensor scale) -> ()");
-  ops.impl("silu_and_mul_quant", torch::kCUDA, &silu_and_mul_quant);
-
-  // Horizontally-fused DeepseekV4-MLA: per-head RMSNorm + GPT-J RoPE for Q, and
-  // GPT-J RoPE + UE8M0 FP8 quant + paged cache insert for KV, all in one
-  // kernel launch. Registered in _C_stable_libtorch (incl. the FlashInfer V4
-  // full-cache bf16/fp8 variants).
-
-  // Quantization ops
-#ifndef USE_ROCM
-
-  // Note about marlin kernel 'workspace' arguments:
-  // Technically these should be mutable since they are modified by the kernel.
-  // But since they are set back to zero once the kernel is finished we can
-  // hand wave and say that they have no net effect.
-  //
-  // The reason to mark 'workspace' as immutable is so that they don't interfere
-  // with using ScalarType arguments in the ops. If they are marked as mutable,
-  // pytorch throws an assert in
-  // 'torch._higher_order_ops._register_effectful_op' that prevents these
-  // kernels from being torch.compile'd.
-  // See the following document for more info on custom types and ops that use
-  // custom types:
-  // https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA
-
 #endif
 }

@@ -792,7 +792,7 @@ RUN --mount=type=cache,target=/opt/uv/cache \
 # Install FlashInfer JIT cache (requires CUDA-version-specific index URL)
 # https://docs.flashinfer.ai/installation.html
 # From versions.json: .flashinfer.version
-ARG FLASHINFER_VERSION=0.6.12
+ARG FLASHINFER_VERSION=0.6.13rc2
 RUN --mount=type=cache,target=/opt/uv/cache \
    uv pip install --system flashinfer-jit-cache==${FLASHINFER_VERSION} \
        --index-url https://flashinfer.ai/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
@@ -75,13 +75,13 @@ RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRO
 # Install UMD
 RUN mkdir neo && \
    cd neo && \
-    wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.24.8/intel-igc-core-2_2.24.8+20344_amd64.deb && \
-    wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.24.8/intel-igc-opencl-2_2.24.8+20344_amd64.deb && \
-    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/intel-ocloc_25.48.36300.8-0_amd64.deb && \
-    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/intel-opencl-icd_25.48.36300.8-0_amd64.deb && \
-    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/libigdgmm12_22.8.2_amd64.deb && \
-    wget https://github.com/intel/compute-runtime/releases/download/25.48.36300.8/libze-intel-gpu1_25.48.36300.8-0_amd64.deb && \
-    wget https://github.com/oneapi-src/level-zero/releases/download/v1.26.0/level-zero_1.26.0+u24.04_amd64.deb && \
+    wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.34.4/intel-igc-core-2_2.34.4+21428_amd64.deb && \
+    wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.34.4/intel-igc-opencl-2_2.34.4+21428_amd64.deb && \
+    wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/intel-ocloc_26.18.38308.1-0_amd64.deb && \
+    wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/intel-opencl-icd_26.18.38308.1-0_amd64.deb && \
+    wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/libigdgmm12_22.10.0_amd64.deb && \
+    wget https://github.com/intel/compute-runtime/releases/download/26.18.38308.1/libze-intel-gpu1_26.18.38308.1-0_amd64.deb && \
+    wget https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero_1.28.2+u24.04_amd64.deb && \
    dpkg -i *.deb && \
    cd .. && \
    rm -rf neo
@@ -68,7 +68,7 @@
      "default": "true"
    },
    "FLASHINFER_VERSION": {
-      "default": "0.6.12"
+      "default": "0.6.13rc2"
    },
    "GDRCOPY_CUDA_VERSION": {
      "default": "12.8"
@@ -133,7 +133,7 @@ The model should inherit protocol `IsAttentionFree` and also implement class met
 For the mamba layers themselves, please use the [`MambaMixer`](../../../vllm/model_executor/layers/mamba/mamba_mixer.py) (for Mamba-1) or [`MambaMixer2`](../../../vllm/model_executor/layers/mamba/mamba_mixer2.py) (for Mamba-2) classes.
 The model should also be added to the `MODELS_CONFIG_MAP` dictionary in [vllm/model_executor/models/config.py](../../../vllm/model_executor/models/config.py) to ensure that the runtime defaults are optimized.

-For case (2), we recommend using as a reference the implementation of [`JambaForCausalLM`](../../../vllm/model_executor/models/jamba.py) (for an example of a model that uses Mamba-1 and attention together) or [`BambaForCausalLM`](../../../vllm/model_executor/models/bamba.py) (for an example of a model that uses Mamba-2 and attention together).
+For case (2), we recommend using as a reference the implementation of [`JambaForCausalLM`](../../../vllm/model_executor/models/jamba.py) (for an example of a model that uses Mamba-1 and attention together) or [`NemotronHForCausalLM`](../../../vllm/model_executor/models/nemotron_h.py) (for an example of a model that uses Mamba-2 and attention together).
 These models should follow the same instructions as case (1), but they should inherit protocol `IsHybrid` (instead of `IsAttentionFree`) and it is *not* necessary to add them to the `MODELS_CONFIG_MAP` (their runtime defaults will be inferred from the protocol).

 For case (3), we recommend looking at the implementation of [`MiniMaxText01ForCausalLM`](../../../vllm/model_executor/models/minimax_text_01.py) or [`Lfm2ForCausalLM`](../../../vllm/model_executor/models/lfm2.py) as a reference, which use custom "mamba-like" layers `MiniMaxText01LinearAttention` and `ShortConv` respectively.
@@ -4,7 +4,7 @@ Deploying vLLM on Kubernetes is a scalable and efficient way to serve machine le

 * **Upstream vLLM compatibility** – It wraps around upstream vLLM without modifying its code.
 * **Ease of use** – Simplified deployment via Helm charts and observability through Grafana dashboards.
-* **High performance** – Optimized for LLM workloads with features like multimodel support, model-aware and prefix-aware routing, fast vLLM bootstrapping, and KV cache offloading with [LMCache](https://github.com/LMCache/LMCache), among others.
+* **High performance** – Optimized for LLM workloads with features like multimodel support, model-aware and prefix-aware routing, fast vLLM bootstrapping, and KV cache offloading with [LMCache](https://github.com/LMCache/LMCache) (wired up in vLLM via `--kv-offloading-backend lmcache`; see the [LMCache examples](https://github.com/vllm-project/vllm/tree/main/examples/disaggregated/lmcache) and [docs.lmcache.ai](https://docs.lmcache.ai)), among others.

 If you are new to Kubernetes, don't worry: in the vLLM production stack [repo](https://github.com/vllm-project/production-stack), we provide a step-by-step [guide](https://github.com/vllm-project/production-stack/blob/main/tutorials/00-install-kubernetes-env.md) and a [short video](https://www.youtube.com/watch?v=EsTJbQtzj0g) to set up everything and get started in **4 minutes**!

@@ -20,7 +20,7 @@ Two main reasons:
 Now supports 9 types of connectors:

 - **ExampleConnector**: refer to [examples/disaggregated/example_connector/run.sh](../../examples/disaggregated/example_connector/run.sh) for the example usage of ExampleConnector disaggregated prefilling.
- **LMCacheConnectorV1**: refer to [examples/disaggregated/lmcache/disagg_prefill_lmcache_v1/disagg_example_nixl.sh](../../examples/disaggregated/lmcache/disagg_prefill_lmcache_v1/disagg_example_nixl.sh) for the example usage of LMCacheConnectorV1 disaggregated prefilling which uses NIXL as the underlying KV transmission.
+- **LMCacheConnectorV1**: refer to [examples/disaggregated/lmcache/disagg_prefill_lmcache_v1/disagg_example_nixl.sh](../../examples/disaggregated/lmcache/disagg_prefill_lmcache_v1/disagg_example_nixl.sh) for the example usage of LMCacheConnectorV1 disaggregated prefilling which uses NIXL as the underlying KV transmission. LMCache also offers a multi-process (MP) mode via `LMCacheMPConnector`, where a standalone `lmcache server` holds the KV cache shared by one or more vLLM instances; see the [LMCache examples](../../examples/disaggregated/lmcache/README.md) and the [LMCache docs](https://docs.lmcache.ai) for setup.
 - **NixlConnector**: refer to [tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh](../../tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh) for the example usage of NixlConnector disaggregated prefilling which support fully async send/recv. For detailed usage guide, see [NixlConnector Usage Guide](nixl_connector_usage.md). For feature compatibility details, see [NixlConnector Compatibility Matrix](nixl_connector_compatibility.md). You may specify one or multiple NIXL transfer backends, such as:

  ```bash
@@ -203,6 +203,7 @@ the vLLM JSON config.
 ### kv_connector_extra_config

 - `load_async` (bool): Enable asynchronous loading for better compute-I/O overlap. Default: `true`.
+- `lookup_async` (bool): Run the external prefix-cache lookup on a background thread so it never blocks the scheduler step. The request is held until the in-flight lookup completes, then resumed on a later step. Default: `false`.
 - `enable_cross_layers_blocks` (bool): Enable cross-layer block packing for reduced store operations. Default: `false`.
 - `lookup_rpc_port` (int): Custom port for the ZMQ lookup RPC socket. Default: `0`.
 - `cache_prefix` (str): Namespace prepended to every store key. Lets separate deployments share one Mooncake master without polluting each other — instances configured with different prefixes never see each other's cached blocks, even for identical prompts. All instances that should share a prefix cache must use the same value. Default: `""` (no prefix; keys are byte-identical to the unprefixed format).
@@ -27,6 +27,7 @@ Currently, there are no pre-built XPU wheels.

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

 ```bash
 git clone https://github.com/vllm-project/vllm.git
@@ -1,10 +1,38 @@
 # LMCache Examples

-This folder demonstrates how to use LMCache for disaggregated prefilling, CPU offloading and KV cache sharing.
+This folder demonstrates how to use LMCache with vLLM v1 for KV cache
+offloading, disaggregated prefilling, and KV cache sharing.

-## 1. Disaggregated Prefill in vLLM v1
+## Integration modes

-This example demonstrates how to run LMCache with disaggregated prefill using NIXL on a single node.
+LMCache integrates with vLLM v1 in two ways:
+
+- **In-process mode** (`LMCacheConnectorV1`): LMCache runs inside the vLLM
+  process and is configured through environment variables or a YAML config
+  file (`LMCACHE_CONFIG_FILE`). This is the simplest way to add single-node
+  CPU/disk offloading.
+- **Multi-process (MP) mode** (`LMCacheMPConnector`): LMCache runs as a
+  standalone server (`lmcache server`) that owns the KV cache storage; one or
+  more vLLM instances connect to it. This is the recommended mode for
+  distributed KV storage and for sharing KV cache across instances. See the
+  [LMCache docs](https://docs.lmcache.ai) for the full MP setup.
+
+## 1. CPU offload (in-process)
+
+- `python cpu_offload_lmcache.py` - CPU offloading with `LMCacheConnectorV1`
+  for vLLM v1.
+
+## 2. CPU offload (multi-process)
+
+- `bash cpu_offload_lmcache_mp.sh` - CPU offloading with `LMCacheMPConnector`,
+  using a standalone `lmcache server`. vLLM provides a built-in shortcut for
+  this setup via `--kv-offloading-backend lmcache` and
+  `--kv-offloading-size <GiB>`.
+
+## 3. Disaggregated Prefill in vLLM v1
+
+This example demonstrates how to run LMCache with disaggregated prefill using
+NIXL on a single node.

 ### Prerequisites

@@ -46,15 +74,7 @@ The main script generates several log files:
 - `decoder.log` - Logs from the decode server
 - `proxy.log` - Logs from the proxy server

-## 2. CPU Offload Examples
+## 4. KV Cache Sharing

- `python cpu_offload_lmcache.py -v v0` - CPU offloading implementation for vLLM v0
- `python cpu_offload_lmcache.py -v v1` - CPU offloading implementation for vLLM v1
-
-## 3. KV Cache Sharing
-
-The `kv_cache_sharing_lmcache_v1.py` example demonstrates how to share KV caches between vLLM v1 instances.
-
-## 4. Disaggregated Prefill in vLLM v0
-
-The `disaggregated_prefill_lmcache_v0.py` provides an example of how to run disaggregated prefill in vLLM v0.
+The `kv_cache_sharing_lmcache_v1.py` example demonstrates how to share KV
+caches between vLLM v1 instances through a centralized LMCache server.
@@ -1,20 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
-This file demonstrates the example usage of cpu offloading
-with LMCache in vLLM v1 or v0.
-
-Usage:
-
-    Specify vLLM version
-
-    -v v0 : Use LMCacheConnector
-            model = mistralai/Mistral-7B-Instruct-v0.2
-            (Includes enable_chunked_prefill = True)
-
-    -v v1 : Use LMCacheConnectorV1 (default)
-            model = meta-llama/Meta-Llama-3.1-8B-Instruct
-            (Without enable_chunked_prefill)
+This file demonstrates the example usage of CPU offloading
+with LMCache in vLLM v1.

 Note that `lmcache` is needed to run this example.
 Requirements:
@@ -23,7 +11,6 @@ Learn more about LMCache environment setup, please refer to:
 https://docs.lmcache.ai/getting_started/installation.html
 """

-import argparse
 import contextlib
 import os
 import time
@@ -39,8 +26,6 @@ from vllm.engine.arg_utils import EngineArgs

 def setup_environment_variables():
    # LMCache-related environment variables
-    # Use experimental features in LMCache
-    os.environ["LMCACHE_USE_EXPERIMENTAL"] = "True"
    # LMCache is set to use 256 tokens per chunk
    os.environ["LMCACHE_CHUNK_SIZE"] = "256"
    # Enable local CPU backend in LMCache
@@ -50,9 +35,9 @@ def setup_environment_variables():


@contextlib.contextmanager
-def build_llm_with_lmcache(lmcache_connector: str, model: str):
+def build_llm_with_lmcache(model: str):
    ktc = KVTransferConfig(
-        kv_connector=lmcache_connector,
+        kv_connector="LMCacheConnectorV1",
        kv_role="kv_both",
    )
    # Set GPU memory utilization to 0.8 for an A40 GPU with 40GB
@@ -92,23 +77,10 @@ def print_output(
    print("-" * 50)


-def parse_args():
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "-v",
-        "--version",
-        choices=["v0", "v1"],
-        default="v1",
-        help="Specify vLLM version (default: v1)",
-    )
-    return parser.parse_args()
-
-
 def main():
-    lmcache_connector = "LMCacheConnectorV1"
    model = "meta-llama/Meta-Llama-3.1-8B-Instruct"
    setup_environment_variables()
-    with build_llm_with_lmcache(lmcache_connector, model) as llm:
+    with build_llm_with_lmcache(model) as llm:
        # This example script runs two requests with a shared prefix.
        # Define the shared prompt and specific prompts
        shared_prompt = "Hello, how are you?" * 1000
@@ -0,0 +1,43 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# CPU offloading with LMCache in multi-process (MP) mode.
+#
+# In MP mode, LMCache runs as a standalone server process (`lmcache server`)
+# that owns the KV cache storage. One or more vLLM instances connect to it via
+# the `LMCacheMPConnector`. This is the recommended way to run LMCache for
+# distributed KV storage and for sharing KV cache across vLLM instances.
+#
+# vLLM ships a built-in shortcut for this setup: pass `--kv-offloading-backend
+# lmcache` together with `--kv-offloading-size <GiB>` and vLLM wires up the
+# `LMCacheMPConnector` for you (it defaults to the LMCache server at
+# tcp://localhost:5555, matching the `lmcache server` default).
+#
+# Requires `lmcache` to be installed (`pip install lmcache`).
+# Learn more: https://docs.lmcache.ai
+set -euo pipefail
+
+MODEL=${MODEL:-meta-llama/Meta-Llama-3.1-8B-Instruct}
+
+# 1. Launch the standalone LMCache server (binds tcp://localhost:5555 by
+#    default). `--l1-size-gb` sets the CPU memory budget for the L1 cache.
+echo "Starting LMCache server..."
+lmcache server --host localhost --port 5555 --l1-size-gb 5 &
+LMCACHE_SERVER_PID=$!
+trap 'kill $LMCACHE_SERVER_PID 2>/dev/null || true' EXIT
+
+# 2. Launch vLLM and offload KV cache to the LMCache server.
+#    The MP connector currently requires the non-hybrid KV cache manager.
+echo "Starting vLLM server with LMCache MP offloading..."
+vllm serve "$MODEL" \
+    --port 8000 \
+    --kv-offloading-size 5 \
+    --kv-offloading-backend lmcache \
+    --disable-hybrid-kv-cache-manager
+
+# Equivalent explicit configuration (instead of the two flags above):
+#   --kv-transfer-config \
+#   '{"kv_connector":"LMCacheMPConnector","kv_role":"kv_both",
+#     "kv_connector_extra_config":{"lmcache.mp.host":"tcp://localhost",
+#                                  "lmcache.mp.port":5555}}'
@@ -1,144 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""
-This file demonstrates the example usage of disaggregated prefilling
-with LMCache.
-We will launch 2 vllm instances (GPU 0 for prefill and GPU 1 for decode),
-and launch an additional LMCache server.
-KV cache is transferred in the following manner:
-vLLM prefill node -> LMCache server -> vLLM decode node.
-
-Note that `pip install lmcache` is needed to run this example.
-Learn more about LMCache in https://github.com/LMCache/LMCache.
-"""
-
-import os
-import subprocess
-import time
-from multiprocessing import Event, Process
-
-from lmcache.experimental.cache_engine import LMCacheEngineBuilder
-from lmcache.integration.vllm.utils import ENGINE_NAME
-
-from vllm import LLM, SamplingParams
-from vllm.config import KVTransferConfig
-
-# LMCache-related environment variables
-# The port to start LMCache server
-port = 8100
-# Use experimental features in LMCache
-os.environ["LMCACHE_USE_EXPERIMENTAL"] = "True"
-# LMCache is set to use 256 tokens per chunk
-os.environ["LMCACHE_CHUNK_SIZE"] = "256"
-# Disable local CPU backend in LMCache
-os.environ["LMCACHE_LOCAL_CPU"] = "False"
-# Set local CPU memory buffer limit to 5.0 GB
-os.environ["LMCACHE_MAX_LOCAL_CPU_SIZE"] = "5.0"
-# Set the remote URL for LMCache server
-os.environ["LMCACHE_REMOTE_URL"] = f"lm://localhost:{port}"
-# Set the serializer/deserializer between vllm and LMCache server
-# `naive` indicates using raw bytes of the tensor without any compression
-os.environ["LMCACHE_REMOTE_SERDE"] = "naive"
-
-prompts = [
-    "Hello, how are you?" * 1000,
-]
-
-
-def run_prefill(prefill_done, prompts):
-    # We use GPU 0 for prefill node.
-    os.environ["CUDA_VISIBLE_DEVICES"] = "0"
-
-    sampling_params = SamplingParams(temperature=0, top_p=0.95, max_tokens=1)
-
-    ktc = KVTransferConfig(
-        kv_connector="LMCacheConnector",
-        kv_role="kv_producer",
-        kv_rank=0,
-        kv_parallel_size=2,
-    )
-    # Set GPU memory utilization to 0.8 for an A40 GPU with 40GB
-    # memory. Reduce the value if your GPU has less memory.
-    llm = LLM(
-        model="mistralai/Mistral-7B-Instruct-v0.2",
-        kv_transfer_config=ktc,
-        max_model_len=8000,
-        gpu_memory_utilization=0.8,
-        enforce_eager=True,
-    )
-
-    # llm.generate(prompts, sampling_params)
-    outputs = llm.generate(prompts, sampling_params)
-    for output in outputs:
-        generated_text = output.outputs[0].text
-        print(f"Generated text: {generated_text!r}")
-    print("Prefill node is finished.")
-    prefill_done.set()
-
-    # Clean up lmcache backend
-    LMCacheEngineBuilder.destroy(ENGINE_NAME)
-
-
-def run_decode(prefill_done, prompts, timeout=1):
-    # We use GPU 1 for decode node.
-    os.environ["CUDA_VISIBLE_DEVICES"] = "1"
-
-    sampling_params = SamplingParams(temperature=0, top_p=0.95, max_tokens=10)
-
-    ktc = KVTransferConfig(
-        kv_connector="LMCacheConnector",
-        kv_role="kv_consumer",
-        kv_rank=1,
-        kv_parallel_size=2,
-    )
-    # Set GPU memory utilization to 0.8 for an A40 GPU with 40GB
-    # of memory. Reduce the value if your GPU has less memory.
-    llm = LLM(
-        model="mistralai/Mistral-7B-Instruct-v0.2",
-        kv_transfer_config=ktc,
-        max_model_len=8000,
-        gpu_memory_utilization=0.8,
-        enforce_eager=True,
-    )
-
-    print("Waiting for prefill node to finish...")
-    prefill_done.wait()
-    time.sleep(timeout)
-
-    outputs = llm.generate(prompts, sampling_params)
-    for output in outputs:
-        generated_text = output.outputs[0].text
-        print(f"Generated text: {generated_text!r}")
-
-    # Clean up lmcache backend
-    LMCacheEngineBuilder.destroy(ENGINE_NAME)
-
-
-def run_lmcache_server(port):
-    server_proc = subprocess.Popen(
-        ["python", "-m", "lmcache.experimental.server", "localhost", str(port)]
-    )
-    return server_proc
-
-
-def main():
-    prefill_done = Event()
-    prefill_process = Process(target=run_prefill, args=(prefill_done, prompts))
-    decode_process = Process(target=run_decode, args=(prefill_done, prompts))
-    lmcache_server_process = run_lmcache_server(port)
-
-    # Start prefill node
-    prefill_process.start()
-
-    # Start decode node
-    decode_process.start()
-
-    # Clean up the processes
-    decode_process.join()
-    prefill_process.terminate()
-    lmcache_server_process.terminate()
-    lmcache_server_process.wait()
-
-
-if __name__ == "__main__":
-    main()
@@ -30,7 +30,6 @@ if [[ $1 == "prefiller" ]]; then

    UCX_TLS=cuda_ipc,cuda_copy,tcp \
        LMCACHE_CONFIG_FILE=$prefill_config_file \
-        LMCACHE_USE_EXPERIMENTAL=True \
        VLLM_ENABLE_V1_MULTIPROCESSING=1 \
        VLLM_WORKER_MULTIPROC_METHOD=spawn \
        CUDA_VISIBLE_DEVICES=0 \
@@ -47,7 +46,6 @@ elif [[ $1 == "decoder" ]]; then

    UCX_TLS=cuda_ipc,cuda_copy,tcp \
        LMCACHE_CONFIG_FILE=$decode_config_file \
-        LMCACHE_USE_EXPERIMENTAL=True \
        VLLM_ENABLE_V1_MULTIPROCESSING=1 \
        VLLM_WORKER_MULTIPROC_METHOD=spawn \
        CUDA_VISIBLE_DEVICES=1 \
@@ -26,8 +26,6 @@ from vllm.config import KVTransferConfig
 # LMCache-related environment variables
 # The port to start LMCache server
 port = 8100
-# Use experimental features in LMCache
-os.environ["LMCACHE_USE_EXPERIMENTAL"] = "True"
 # LMCache is set to use 256 tokens per chunk
 os.environ["LMCACHE_CHUNK_SIZE"] = "256"
 # Disable local CPU backend in LMCache
@@ -11,13 +11,14 @@ transformers >= 5.5.3
 tokenizers >= 0.21.1  # Required for fast incremental detokenization.
 safetensors >= 0.6.2  # MXFP4/MXFP6 dtype support (F8_E8M0, F4) added in 0.6.0: https://github.com/huggingface/safetensors/pull/611
 protobuf >= 5.29.6, !=6.30.*, !=6.31.*, !=6.32.*, !=6.33.0.*, !=6.33.1.*, !=6.33.2.*, !=6.33.3.*, !=6.33.4.* # Required by LlamaTokenizer, gRPC. CVE-2026-0994
-fastapi[standard] >= 0.115.0 # Required by FastAPI's form models in the OpenAI API server's audio transcriptions endpoint.
+fastapi[standard] >= 0.133.0, < 0.137.0 # First version supporting Starlette 1.0; < 0.137.0 avoids route-tree change that breaks model-hosting-container-standards handler overrides.
+starlette >= 1.0.1 # CVE-2026-48710: Host header injection in < 1.0.1
 aiohttp >= 3.13.3
 openai >= 2.0.0  # For Responses API with reasoning content
 pydantic >= 2.12.0
 prometheus_client >= 0.18.0
 pillow  # Required for image processing
-prometheus-fastapi-instrumentator >= 7.0.0
+prometheus-fastapi-instrumentator >= 8.0.0 # v8 unblocks starlette >= 1.0
 tiktoken >= 0.6.0  # Required for DBRX tokenizer
 lm-format-enforcer == 0.11.3
 llguidance >= 1.7.0, < 1.8.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64" or platform_machine == "ppc64le"
@@ -9,8 +9,8 @@ torchaudio==2.11.0
 # These must be updated alongside torch
 torchvision==0.26.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
 # FlashInfer should be updated together with the Dockerfile
-flashinfer-python==0.6.12
-flashinfer-cubin==0.6.12
+flashinfer-python==0.6.13rc2
+flashinfer-cubin==0.6.13rc2
 apache-tvm-ffi==0.1.9
 tilelang==0.1.9
 nvidia-cudnn-frontend>=1.19.1
@@ -40,7 +40,7 @@ lm-eval[api]>=0.4.12 # required for model evaluation test
 mteb[bm25s]>=2, <3 # required for mteb test
 transformers==5.5.3
 tokenizers==0.22.2
-schemathesis>=3.39.15 # Required for openai schema test.
+schemathesis>=4.0.0 # Required for openai schema test.
 # quantization
 bitsandbytes==0.49.2
 buildkite-test-collector==0.1.9
@@ -35,14 +35,11 @@ arctic-inference==0.1.1
    # via -r requirements/test/cuda.in
 argcomplete==3.5.1
    # via datamodel-code-generator
-arrow==1.3.0
-    # via isoduration
 attrs==24.2.0
    # via
    #   aiohttp
    #   hypothesis
    #   jsonschema
-    #   pytest-subtests
    #   referencing
 audioread==3.0.1
    # via librosa
@@ -57,9 +54,7 @@ azure-identity==1.25.2
 azure-storage-blob==12.28.0
    # via runai-model-streamer-azure
 backoff==2.2.1
-    # via
-    #   -r requirements/test/cuda.in
-    #   schemathesis
+    # via -r requirements/test/cuda.in
 bitsandbytes==0.49.2
    # via -r requirements/test/cuda.in
 black==24.10.0
@@ -110,7 +105,6 @@ colorama==0.4.6
    # via
    #   perceptron
    #   sacrebleu
-    #   schemathesis
 colorful==0.5.6
    # via ray
 colorlog==6.10.1
@@ -183,7 +177,7 @@ et-xmlfile==2.0.0
    # via openpyxl
 evaluate==0.4.3
    # via lm-eval
-fastapi==0.128.0
+fastapi==0.136.3
    # via
    #   -c requirements/common.txt
    #   gpt-oss
@@ -206,8 +200,6 @@ filelock==3.16.1
    #   virtualenv
 fonttools==4.55.0
    # via matplotlib
-fqdn==1.5.1
-    # via jsonschema
 frozendict==2.4.6
    # via einx
 frozenlist==1.5.0
@@ -269,7 +261,7 @@ h11==0.14.0
    #   uvicorn
 h2==4.3.0
    # via httpx
-harfile==0.3.0
+harfile==0.5.0
    # via schemathesis
 hf-xet==1.4.3
    # via huggingface-hub
@@ -309,7 +301,7 @@ hypothesis==6.131.0
    #   hypothesis-graphql
    #   hypothesis-jsonschema
    #   schemathesis
-hypothesis-graphql==0.11.1
+hypothesis-graphql==0.13.0
    # via schemathesis
 hypothesis-jsonschema==0.23.1
    # via schemathesis
@@ -318,7 +310,6 @@ idna==3.10
    #   anyio
    #   email-validator
    #   httpx
-    #   jsonschema
    #   requests
    #   yarl
 imagehash==4.3.2
@@ -335,8 +326,6 @@ instanttensor==0.1.5
    # via -r requirements/test/cuda.in
 isodate==0.7.2
    # via azure-storage-blob
-isoduration==20.11.0
-    # via jsonschema
 isort==5.13.2
    # via datamodel-code-generator
 jinja2==3.1.6
@@ -356,15 +345,14 @@ joblib==1.4.2
    #   librosa
    #   nltk
    #   scikit-learn
-jsonpointer==3.0.0
-    # via jsonschema
 jsonschema==4.23.0
    # via
    #   -c requirements/common.txt
    #   hypothesis-jsonschema
    #   mistral-common
    #   ray
-    #   schemathesis
+jsonschema-rs==0.46.5
+    # via schemathesis
 jsonschema-specifications==2024.10.1
    # via jsonschema
 junit-xml==1.9
@@ -715,18 +703,20 @@ pydantic-core==2.41.1
 pydantic-extra-types==2.10.5
    # via mistral-common
 pygments==2.18.0
-    # via rich
+    # via
+    #   pytest
+    #   rich
 pyjwt==2.11.0
    # via msal
 pyparsing==3.2.0
    # via matplotlib
-pyrate-limiter==3.7.0
+pyrate-limiter==4.4.0
    # via schemathesis
 pystemmer==3.0.0
    # via mteb
 pytablewriter==1.2.0
    # via lm-eval
-pytest==8.3.5
+pytest==9.1.0
    # via
    #   -r requirements/test/cuda.in
    #   buildkite-test-collector
@@ -737,10 +727,9 @@ pytest==8.3.5
    #   pytest-mock
    #   pytest-rerunfailures
    #   pytest-shard
-    #   pytest-subtests
    #   pytest-timeout
    #   schemathesis
-pytest-asyncio==0.24.0
+pytest-asyncio==1.4.0
    # via -r requirements/test/cuda.in
 pytest-cov==6.3.0
    # via -r requirements/test/cuda.in
@@ -752,13 +741,10 @@ pytest-rerunfailures==14.0
    # via -r requirements/test/cuda.in
 pytest-shard==0.1.2
    # via -r requirements/test/cuda.in
-pytest-subtests==0.14.1
-    # via schemathesis
 pytest-timeout==2.3.1
    # via -r requirements/test/cuda.in
 python-dateutil==2.9.0.post0
    # via
-    #   arrow
    #   botocore
    #   matplotlib
    #   pandas
@@ -829,15 +815,12 @@ requests==2.32.3
    #   tiktoken
 responses==0.25.3
    # via genai-perf
-rfc3339-validator==0.1.4
-    # via jsonschema
-rfc3987==1.3.8
-    # via jsonschema
 rich==13.9.4
    # via
    #   genai-perf
    #   mteb
    #   perceptron
+    #   schemathesis
    #   typer
 rouge-score==0.1.2
    # via lm-eval
@@ -868,7 +851,7 @@ safetensors==0.7.0
    #   segmentation-models-pytorch
    #   timm
    #   transformers
-schemathesis==3.39.15
+schemathesis==4.21.6
    # via -r requirements/test/cuda.in
 scikit-image==0.25.2
    # via albumentations
@@ -912,7 +895,6 @@ six==1.16.0
    #   junit-xml
    #   opencensus
    #   python-dateutil
-    #   rfc3339-validator
    #   rouge-score
 smart-open==7.1.0
    # via ray
@@ -938,10 +920,10 @@ sqlalchemy==2.0.41
    #   optuna
 sqlitedict==2.1.0
    # via lm-eval
-starlette==0.50.0
+starlette==1.3.1
    # via
+    #   -c requirements/common.txt
    #   fastapi
-    #   schemathesis
    #   starlette-testclient
 starlette-testclient==0.4.1
    # via schemathesis
@@ -966,6 +948,7 @@ tenacity==9.1.2
    #   gpt-oss
    #   lm-eval
    #   plotly
+    #   schemathesis
 tensorizer==2.10.1
    # via -r requirements/test/cuda.in
 termcolor==3.1.0
@@ -990,10 +973,6 @@ tokenizers==0.22.2
    #   -c requirements/common.txt
    #   -r requirements/test/cuda.in
    #   transformers
-tomli==2.2.1
-    # via schemathesis
-tomli-w==1.2.0
-    # via schemathesis
 torch==2.11.0+cu130
    # via
    #   -c requirements/cuda.txt
@@ -1066,8 +1045,6 @@ typer==0.15.2
    #   huggingface-hub
    #   perceptron
    #   transformers
-types-python-dateutil==2.9.0.20241206
-    # via arrow
 typing-extensions==4.15.0
    # via
    #   -c requirements/common.txt
@@ -1092,6 +1069,8 @@ typing-extensions==4.15.0
    #   pydantic
    #   pydantic-core
    #   pydantic-extra-types
+    #   pytest-asyncio
+    #   schemathesis
    #   sentence-transformers
    #   sqlalchemy
    #   starlette
@@ -1099,11 +1078,11 @@ typing-extensions==4.15.0
    #   typer
    #   typing-inspection
 typing-inspection==0.4.2
-    # via pydantic
+    # via
+    #   fastapi
+    #   pydantic
 tzdata==2024.2
    # via pandas
-uri-template==1.3.0
-    # via jsonschema
 urllib3==2.2.3
    # via
    #   blobfile
@@ -1122,8 +1101,6 @@ vocos==0.1.0
    # via -r requirements/test/cuda.in
 wcwidth==0.2.13
    # via ftfy
-webcolors==24.11.1
-    # via jsonschema
 werkzeug==3.1.3
    # via schemathesis
 word2number==1.1
@@ -1135,8 +1112,6 @@ xxhash==3.5.0
    #   datasets
    #   evaluate
 yarl==1.17.1
-    # via
-    #   aiohttp
-    #   schemathesis
+    # via aiohttp
 zipp==3.23.0
    # via importlib-metadata
@@ -31,7 +31,7 @@ lm-eval[api]>=0.4.12 # required for model evaluation test
 mteb[bm25s]>=2, <3 # required for mteb test
 transformers==5.5.3
 tokenizers==0.22.2
-schemathesis>=3.39.15 # Required for openai schema test.
+schemathesis>=4.0.0 # Required for openai schema test.
 # quantization
 bitsandbytes>=0.49.2
 buildkite-test-collector==0.1.9
@@ -39,7 +39,7 @@ lm-eval[api]>=0.4.12 # required for model evaluation test
 mteb[bm25s]>=2, <3 # required for mteb test
 transformers==5.5.3
 tokenizers==0.22.2
-schemathesis>=3.39.15 # Required for openai schema test
+schemathesis>=4.0.0 # Required for openai schema test
 # quantization
 bitsandbytes==0.49.2
 buildkite-test-collector==0.1.9
@@ -51,15 +51,12 @@ arctic-inference==0.1.1
    # via -r requirements/test/rocm.in
 argcomplete==3.6.3
    # via datamodel-code-generator
-arrow==1.4.0
-    # via isoduration
 astor==0.8.1
    # via depyf
 attrs==26.1.0
    # via
    #   aiohttp
    #   jsonschema
-    #   pytest-subtests
    #   referencing
 audioread==3.0.1
    # via librosa
@@ -74,9 +71,7 @@ azure-identity==1.25.3
 azure-storage-blob==12.28.0
    # via runai-model-streamer-azure
 backoff==2.2.1
-    # via
-    #   -r requirements/test/rocm.in
-    #   schemathesis
+    # via -r requirements/test/rocm.in
 bitsandbytes==0.49.2
    # via -r requirements/test/rocm.in
 black==26.3.1
@@ -139,7 +134,6 @@ colorama==0.4.6
    # via
    #   perceptron
    #   sacrebleu
-    #   schemathesis
 colorful==0.5.8
    # via ray
 colorlog==6.10.1
@@ -258,8 +252,6 @@ filelock==3.25.2
    #   virtualenv
 fonttools==4.62.1
    # via matplotlib
-fqdn==1.5.1
-    # via jsonschema
 frozendict==2.4.7
    # via einx
 frozenlist==1.8.0
@@ -328,7 +320,7 @@ h11==0.16.0
    #   uvicorn
 h2==4.3.0
    # via httpx
-harfile==0.4.0
+harfile==0.5.0
    # via schemathesis
 hf-xet==1.4.3
    # via huggingface-hub
@@ -378,7 +370,7 @@ hypothesis==6.151.9
    #   hypothesis-graphql
    #   hypothesis-jsonschema
    #   schemathesis
-hypothesis-graphql==0.12.0
+hypothesis-graphql==0.13.0
    # via schemathesis
 hypothesis-jsonschema==0.23.1
    # via schemathesis
@@ -387,7 +379,6 @@ idna==3.11
    #   anyio
    #   email-validator
    #   httpx
-    #   jsonschema
    #   requests
    #   yarl
 ijson==3.5.0
@@ -408,8 +399,6 @@ interegular==0.3.3
    # via lm-format-enforcer
 isodate==0.7.2
    # via azure-storage-blob
-isoduration==20.11.0
-    # via jsonschema
 isort==8.0.1
    # via datamodel-code-generator
 jinja2==3.1.6
@@ -435,8 +424,6 @@ joblib==1.5.3
    #   librosa
    #   nltk
    #   scikit-learn
-jsonpointer==3.1.0
-    # via jsonschema
 jsonschema==4.26.0
    # via
    #   -c requirements/common.txt
@@ -445,7 +432,8 @@ jsonschema==4.26.0
    #   mcp
    #   mistral-common
    #   ray
-    #   schemathesis
+jsonschema-rs==0.46.5
+    # via schemathesis
 jsonschema-specifications==2025.9.1
    # via jsonschema
 junit-xml==1.9
@@ -792,7 +780,7 @@ prometheus-client==0.24.1
    #   opentelemetry-exporter-prometheus
    #   prometheus-fastapi-instrumentator
    #   ray
-prometheus-fastapi-instrumentator==7.1.0
+prometheus-fastapi-instrumentator==8.0.0
    # via
    #   -c requirements/common.txt
    #   -r requirements/test/../common.txt
@@ -876,20 +864,22 @@ pydantic-settings==2.13.1
    #   fastapi
    #   mcp
 pygments==2.19.2
-    # via rich
+    # via
+    #   pytest
+    #   rich
 pyjwt==2.12.1
    # via
    #   mcp
    #   msal
 pyparsing==3.3.2
    # via matplotlib
-pyrate-limiter==3.9.0
+pyrate-limiter==4.4.0
    # via schemathesis
 pystemmer==3.0.0
    # via mteb
 pytablewriter==1.2.1
    # via lm-eval
-pytest==8.3.5
+pytest==9.1.0
    # via
    #   -r requirements/test/rocm.in
    #   buildkite-test-collector
@@ -900,10 +890,9 @@ pytest==8.3.5
    #   pytest-mock
    #   pytest-rerunfailures
    #   pytest-shard
-    #   pytest-subtests
    #   pytest-timeout
    #   schemathesis
-pytest-asyncio==0.24.0
+pytest-asyncio==1.4.0
    # via -r requirements/test/rocm.in
 pytest-cov==6.3.0
    # via -r requirements/test/rocm.in
@@ -915,13 +904,10 @@ pytest-rerunfailures==14.0
    # via -r requirements/test/rocm.in
 pytest-shard==0.1.2
    # via -r requirements/test/rocm.in
-pytest-subtests==0.14.2
-    # via schemathesis
 pytest-timeout==2.3.1
    # via -r requirements/test/rocm.in
 python-dateutil==2.9.0.post0
    # via
-    #   arrow
    #   botocore
    #   matplotlib
    #   pandas
@@ -1016,16 +1002,13 @@ requests==2.32.5
    #   tiktoken
 responses==0.26.0
    # via genai-perf
-rfc3339-validator==0.1.4
-    # via jsonschema
-rfc3987==1.3.8
-    # via jsonschema
 rich==14.3.3
    # via
    #   genai-perf
    #   mteb
    #   perceptron
    #   rich-toolkit
+    #   schemathesis
    #   typer
 rich-toolkit==0.19.7
    # via
@@ -1063,7 +1046,7 @@ safetensors==0.7.0
    #   segmentation-models-pytorch
    #   timm
    #   transformers
-schemathesis==3.39.15
+schemathesis==4.21.6
    # via -r requirements/test/rocm.in
 scikit-image==0.26.0
    # via albumentations
@@ -1120,7 +1103,6 @@ six==1.17.0
    #   junit-xml
    #   opencensus
    #   python-dateutil
-    #   rfc3339-validator
    #   rouge-score
 smart-open==7.5.1
    # via ray
@@ -1149,13 +1131,14 @@ sqlitedict==2.1.0
    # via lm-eval
 sse-starlette==3.3.4
    # via mcp
-starlette==0.52.1
+starlette==1.3.1
    # via
+    #   -c requirements/common.txt
+    #   -r requirements/test/../common.txt
    #   fastapi
    #   mcp
    #   model-hosting-container-standards
    #   prometheus-fastapi-instrumentator
-    #   schemathesis
    #   sse-starlette
    #   starlette-testclient
 starlette-testclient==0.4.1
@@ -1182,6 +1165,7 @@ tenacity==9.1.4
    # via
    #   gpt-oss
    #   lm-eval
+    #   schemathesis
 tensorizer==2.10.1
    # via
    #   -c requirements/rocm.txt
@@ -1215,10 +1199,6 @@ tokenizers==0.22.2
    #   -r requirements/test/../common.txt
    #   -r requirements/test/rocm.in
    #   transformers
-tomli==2.4.0
-    # via schemathesis
-tomli-w==1.2.0
-    # via schemathesis
 torch-c-dlpack-ext==0.1.5
    # via tilelang
 tqdm==4.67.3
@@ -1301,8 +1281,10 @@ typing-extensions==4.15.0
    #   pydantic
    #   pydantic-core
    #   pydantic-extra-types
+    #   pytest-asyncio
    #   referencing
    #   rich-toolkit
+    #   schemathesis
    #   sentence-transformers
    #   sqlalchemy
    #   starlette
@@ -1317,10 +1299,6 @@ typing-inspection==0.4.2
    #   mcp
    #   pydantic
    #   pydantic-settings
-tzdata==2025.3
-    # via arrow
-uri-template==1.3.0
-    # via jsonschema
 urllib3==2.6.3
    # via
    #   blobfile
@@ -1351,8 +1329,6 @@ watchfiles==1.1.1
    #   uvicorn
 wcwidth==0.6.0
    # via ftfy
-webcolors==25.10.0
-    # via jsonschema
 websockets==16.0
    # via uvicorn
 werkzeug==3.1.6
@@ -1370,9 +1346,7 @@ xxhash==3.6.0
    #   datasets
    #   evaluate
 yarl==1.23.0
-    # via
-    #   aiohttp
-    #   schemathesis
+    # via aiohttp
 z3-solver==4.15.4.0
    # via tilelang
 zipp==3.23.0
@@ -593,8 +593,9 @@ soxr==0.5.0.post1
    #   mistral-common
 sqlitedict==2.1.0
    # via lm-eval
-starlette==1.0.0
+starlette==1.3.1
    # via
+    #   -c requirements/common.txt
    #   fastapi
    #   starlette-testclient
 starlette-testclient==0.4.1
@@ -17,4 +17,4 @@ torchaudio
 torchvision

 auto_round_lib>=0.13.3
-vllm_xpu_kernels @ https://github.com/vllm-project/vllm-xpu-kernels/releases/download/v0.1.9.1/vllm_xpu_kernels-0.1.9.1-cp38-abi3-manylinux_2_28_x86_64.whl
+vllm_xpu_kernels @ https://github.com/vllm-project/vllm-xpu-kernels/releases/download/v0.1.10/vllm_xpu_kernels-0.1.10-cp38-abi3-manylinux_2_28_x86_64.whl
@@ -769,6 +769,7 @@ class precompiled_wheel_utils:
                            "vllm/_C.abi3.so",
                            "vllm/_C_stable_libtorch.abi3.so",
                            "vllm/_moe_C_stable_libtorch.abi3.so",
+                            "vllm/_qutlass_C.abi3.so",
                            "vllm/_flashmla_C.abi3.so",
                            "vllm/_flashmla_extension_C.abi3.so",
                            "vllm/_sparse_flashmla_C.abi3.so",
@@ -1135,6 +1136,7 @@ if _is_cuda():
        # DeepGEMM requires CUDA 12.3+ (SM90/SM100)
        # Optional since it won't build on unsupported architectures
        ext_modules.append(CMakeExtension(name="vllm._deep_gemm_C", optional=True))
+        ext_modules.append(CMakeExtension(name="vllm._qutlass_C", optional=True))
    # fmha_sm100 is a Python/CuTe-DSL package installed into vllm.third_party.
    ext_modules.append(CMakeExtension(name="vllm.fmha_sm100", optional=True))

@@ -1149,7 +1151,8 @@ if _is_cpu():
        ext_modules.append(CMakeExtension(name="vllm._C"))

 if _build_custom_ops():
-    ext_modules.append(CMakeExtension(name="vllm._C"))
+    if _is_hip():
+        ext_modules.append(CMakeExtension(name="vllm._C"))
    if _is_cuda() or _is_hip():
        ext_modules.append(CMakeExtension(name="vllm._C_stable_libtorch"))
        ext_modules.append(CMakeExtension(name="vllm._moe_C_stable_libtorch"))
@@ -1168,7 +1171,15 @@ package_data = {
        "third_party/deep_gemm/include/**/*.h",
        "third_party/deep_gemm/include/**/*.hpp",
        # fmha_sm100 sparse CuTe-DSL helper kernels (vendored via cmake)
+        "third_party/fmha_sm100/csrc/**/*.cu",
+        "third_party/fmha_sm100/csrc/**/*.h",
+        "third_party/fmha_sm100/csrc/**/*.jinja",
+        "third_party/fmha_sm100/csrc/**/*.cu.jinja",
        "third_party/fmha_sm100/cute/**/*.cu",
+        "third_party/fmha_sm100/cutlass/include/**/*.h",
+        "third_party/fmha_sm100/cutlass/include/**/*.hpp",
+        "third_party/fmha_sm100/cutlass/tools/util/include/**/*.h",
+        "third_party/fmha_sm100/cutlass/tools/util/include/**/*.hpp",
    ]
 }

@@ -649,3 +649,196 @@ def test_cloud_storage_tokenizer_skips_get_model_path(monkeypatch):
    args = EngineArgs(model="s3://bucket/model", tokenizer="s3://bucket/tokenizer")
    assert args.model == "s3://bucket/model"
    assert args.tokenizer == "s3://bucket/tokenizer"
+
+
+class TestDeviceIds:
+    def test_device_ids_with_cvd_out_of_range(self, monkeypatch):
+        """--device-ids index beyond the CVD set raises ValueError."""
+        from vllm.platforms import current_platform
+
+        key = current_platform.device_control_env_var
+        monkeypatch.setenv(key, "4,5")
+        args = EngineArgs(model="m", device_ids=[0, 2])
+        with pytest.raises(ValueError, match="out of range"):
+            args._resolve_device_ids()
+
+    def test_device_ids_with_cvd_resolve_to_physical_ids(self, monkeypatch):
+        """--device-ids are CVD-local indices resolved to physical ids."""
+        from vllm.platforms import current_platform
+
+        key = current_platform.device_control_env_var
+        monkeypatch.setenv(key, "4,5")
+        args = EngineArgs(model="m", device_ids=[0, 1])
+        assert args._resolve_device_ids() == [4, 5]
+
+    def test_device_ids_with_uuid_cvd_resolve_to_physical_ids(self, monkeypatch):
+        """--device-ids support UUID CVD values resolved by the platform."""
+        from vllm.platforms import current_platform
+
+        key = current_platform.device_control_env_var
+        monkeypatch.setenv(key, "GPU-abcd1234,GPU-ef567890")
+        monkeypatch.setattr(
+            type(current_platform),
+            "device_control_id_to_physical_device_id",
+            classmethod(
+                lambda cls, device_id: {"GPU-abcd1234": 4, "GPU-ef567890": 5}[device_id]
+            ),
+        )
+
+        args = EngineArgs(model="m", device_ids=[0, 1])
+        assert args._resolve_device_ids() == [4, 5]
+
+    def test_device_ids_with_uuid_args_resolve_to_physical_ids(self, monkeypatch):
+        """UUID --device-ids are resolved to physical IDs immediately."""
+        from vllm.platforms import current_platform
+
+        monkeypatch.setattr(
+            type(current_platform),
+            "device_control_id_to_physical_device_id",
+            classmethod(lambda cls, device_id: {"GPU-abcd1234": 4}[device_id]),
+        )
+
+        args = EngineArgs(model="m", device_ids=["GPU-abcd1234"])
+        assert args._resolve_device_ids() == [4]
+
+    def test_device_ids_reject_mixed_integer_and_uuid_args(self):
+        """--device-ids must not mix CVD indices and UUIDs."""
+        args = EngineArgs(model="m", device_ids=[0, "GPU-abcd1234"])
+        with pytest.raises(ValueError, match="must not mix"):
+            args._resolve_device_ids()
+
+    def test_no_device_ids(self):
+        """No --device-ids returns None."""
+        args = EngineArgs(model="m")
+        assert args._resolve_device_ids() is None
+
+    def test_cli_parsing(self):
+        """--device-ids parses comma-separated string from CLI."""
+        parser = FlexibleArgumentParser()
+        EngineArgs.add_cli_args(parser)
+        parsed = parser.parse_args(["--model", "m", "--device-ids", "0,2,4"])
+        assert parsed.device_ids == [0, 2, 4]
+
+    def test_cli_parsing_uuid(self):
+        """--device-ids parses comma-separated UUID strings from CLI."""
+        parser = FlexibleArgumentParser()
+        EngineArgs.add_cli_args(parser)
+        parsed = parser.parse_args(
+            ["--model", "m", "--device-ids", "GPU-abcd1234,GPU-ef567890"]
+        )
+        assert parsed.device_ids == ["GPU-abcd1234", "GPU-ef567890"]
+
+    def test_assigned_physical_gpu_ids_are_physical_with_cvd(self, monkeypatch):
+        """assigned_physical_gpu_ids are already physical and not composed with CVD."""
+        import vllm.platforms.interface as platform_interface
+        from vllm.platforms import current_platform
+
+        monkeypatch.setattr(platform_interface, "_assigned_physical_gpu_ids", [4, 5])
+        monkeypatch.setenv(current_platform.device_control_env_var, "4,5")
+
+        assert current_platform.device_id_to_physical_device_id(0) == 4
+        assert current_platform.device_id_to_physical_device_id(1) == 5
+        assert current_platform.logical_device_id_to_visible_device_id(0) == 0
+        assert current_platform.logical_device_id_to_visible_device_id(1) == 1
+
+    def test_assigned_physical_gpu_ids_map_to_visible_uuid_cvd(self, monkeypatch):
+        """Physical IDs map back to visible ordinals when CVD uses UUIDs."""
+        import vllm.platforms.interface as platform_interface
+        from vllm.platforms import current_platform
+
+        monkeypatch.setattr(platform_interface, "_assigned_physical_gpu_ids", [5])
+        monkeypatch.setenv(
+            current_platform.device_control_env_var,
+            "GPU-abcd1234,GPU-ef567890",
+        )
+        monkeypatch.setattr(
+            type(current_platform),
+            "device_control_id_to_physical_device_id",
+            classmethod(
+                lambda cls, device_id: {"GPU-abcd1234": 4, "GPU-ef567890": 5}[device_id]
+            ),
+        )
+
+        assert current_platform.logical_device_id_to_visible_device_id(0) == 1
+
+    def test_device_ids_reject_duplicates(self):
+        """--device-ids must not contain duplicate entries."""
+        args = EngineArgs(model="m", device_ids=[2, 2])
+        with pytest.raises(ValueError, match="duplicates"):
+            args._resolve_device_ids()
+
+    def test_cli_parsing_strips_whitespace(self):
+        """--device-ids tolerates whitespace around commas."""
+        parser = FlexibleArgumentParser()
+        EngineArgs.add_cli_args(parser)
+        parsed = parser.parse_args(["--model", "m", "--device-ids", "0, 2, 4"])
+        assert parsed.device_ids == [0, 2, 4]
+
+    def test_visible_ordinal_to_physical_ignores_assigned_ids(self, monkeypatch):
+        """visible_device_id_to_physical_device_id maps torch device ordinals,
+        independent of the logical-to-physical mapping.
+
+        Regression test: CustomAllreduce passes device.index (a visible
+        ordinal) and must not index into assigned_physical_gpu_ids, which
+        raised IndexError for non-identity --device-ids like [2, 3].
+        """
+        import vllm.platforms.interface as platform_interface
+        from vllm.platforms import current_platform
+
+        monkeypatch.setattr(platform_interface, "_assigned_physical_gpu_ids", [2, 3])
+        monkeypatch.delenv(current_platform.device_control_env_var, raising=False)
+
+        # CVD unset: visible ordinal == physical ID, even beyond the
+        # assigned list's length.
+        assert current_platform.visible_device_id_to_physical_device_id(2) == 2
+        assert current_platform.visible_device_id_to_physical_device_id(3) == 3
+
+        monkeypatch.setenv(current_platform.device_control_env_var, "4,5")
+        assert current_platform.visible_device_id_to_physical_device_id(1) == 5
+        with pytest.raises(IndexError, match="out of range"):
+            current_platform.visible_device_id_to_physical_device_id(2)
+
+
+class TestDpDeviceIdSharding:
+    def test_dp_supervisor_device_ids_stay_env_relative(self):
+        """Regression test: the DP supervisor must pass env-relative indices,
+        not physical IDs, because each child re-resolves --device-ids
+        against its inherited device-control env var."""
+        import argparse
+
+        from vllm.entrypoints.openai.dp_supervisor import _build_device_ids
+
+        args = argparse.Namespace(
+            tensor_parallel_size=2, pipeline_parallel_size=1, device_ids=None
+        )
+        assert _build_device_ids(args, local_rank=0) == [0, 1]
+        assert _build_device_ids(args, local_rank=1) == [2, 3]
+
+    def test_dp_supervisor_shards_user_device_ids(self):
+        """User-provided --device-ids are sharded across DP children."""
+        import argparse
+
+        from vllm.entrypoints.openai.dp_supervisor import _build_device_ids
+
+        args = argparse.Namespace(
+            tensor_parallel_size=2, pipeline_parallel_size=1, device_ids=[4, 5, 6, 7]
+        )
+        assert _build_device_ids(args, local_rank=0) == [4, 5]
+        assert _build_device_ids(args, local_rank=1) == [6, 7]
+        with pytest.raises(ValueError, match="needs devices"):
+            _build_device_ids(args, local_rank=2)
+
+    def test_dp_rank_shards_user_assigned_gpu_ids(self):
+        """get_physical_gpu_ids_for_local_dp_rank slices the user-provided
+        --device-ids list instead of recomputing from the env var."""
+        from vllm.platforms import current_platform
+        from vllm.v1.engine.utils import get_physical_gpu_ids_for_local_dp_rank
+
+        evar = current_platform.device_control_env_var
+        assert get_physical_gpu_ids_for_local_dp_rank(
+            evar, local_dp_rank=1, world_size=2, user_assigned_gpu_ids=[4, 5, 6, 7]
+        ) == [6, 7]
+        with pytest.raises(ValueError, match="needs devices"):
+            get_physical_gpu_ids_for_local_dp_rank(
+                evar, local_dp_rank=2, world_size=2, user_assigned_gpu_ids=[4, 5, 6, 7]
+            )
@@ -1096,3 +1096,50 @@ class TestMessageStartIncludesTypeAndRole:
        message = events[0][1]["message"]
        assert message["type"] == "message"
        assert message["role"] == "assistant"
+
+
+# ======================================================================
+# Auto-detection of system-first template requirement
+# ======================================================================
+
+
+Q35_TEMPLATE = (
+    "{%- for message in messages %}"
+    "{%- if message.role == 'system' %}"
+    "{%- if not loop.first %}"
+    "{{- raise_exception('System message must be at the beginning.') }}"
+    "{%- endif %}"
+    "{%- endif %}"
+    "{%- endfor %}"
+)
+
+
+class TestDetectMergeInlineSystem:
+    """Verify _detect_merge_inline_system auto-detection.
+
+    Tests three scenarios:
+    1. Template with system-first guard (e.g. Qwen) → merge needed
+    2. Template without restrictions → no merge, cache-friendly
+    3. No template provided → safe default: merge
+    """
+
+    def test_qwen_template_requires_merge(self):
+        """Template with loop.first guard rejects mid-conversation system."""
+        assert (
+            AnthropicServingMessages._detect_merge_inline_system(Q35_TEMPLATE) is True
+        )
+
+    def test_no_restriction_no_merge(self):
+        """Template without restriction accepts mid-conversation system."""
+        assert (
+            AnthropicServingMessages._detect_merge_inline_system(
+                "{%- for message in messages %}"
+                "{{- message.role }}: {{ message.content }}\n"
+                "{%- endfor %}"
+            )
+            is False
+        )
+
+    def test_no_template_defaults_merge(self):
+        """No chat_template → conservative default: merge."""
+        assert AnthropicServingMessages._detect_merge_inline_system(None) is True
@@ -364,7 +364,7 @@ class MockVLLMServer:
        await self._serve_task


-def launch_mock_vllm(child_args: argparse.Namespace, env_updates: dict[str, str]):
+def launch_mock_vllm(child_args: argparse.Namespace):
    logger.info("Launching mock vLLM on port %s", child_args.port)
    mock_vllm = MockVLLMServer(
        port=child_args.port,
@@ -375,7 +375,7 @@ def launch_mock_vllm(child_args: argparse.Namespace, env_updates: dict[str, str]


 def launch_mock_vllm_with_drain(
-    child_args: argparse.Namespace, env_updates: dict[str, str]
+    child_args: argparse.Namespace,
 ):
    logger.info("Launching mock vLLM with 15s drain on port %s", child_args.port)
    mock_vllm = MockVLLMServer(
@@ -6,15 +6,22 @@ from typing import Final
 import pytest
 import schemathesis
 from hypothesis import HealthCheck, settings
-from schemathesis import GenerationConfig
-from schemathesis.models import Case
+from schemathesis import GenerationMode
+from schemathesis.config import (
+    ChecksConfig,
+    CoveragePhaseConfig,
+    GenerationConfig,
+    PhasesConfig,
+    PositiveDataAcceptanceConfig,
+    ProjectConfig,
+    ProjectsConfig,
+    SchemathesisConfig,
+)

 from vllm.platforms import current_platform

 from ...utils import RemoteOpenAIServer

-schemathesis.experimental.OPEN_API_3_1.enable()
-
 MODEL_NAME = "HuggingFaceTB/SmolVLM-256M-Instruct"
 MAXIMUM_IMAGES = 2
 _ROCM_TIMEOUT_MULTIPLIER = 3 if current_platform.is_rocm() else 1
@@ -44,21 +51,38 @@ def server():
@pytest.fixture(scope="module")
 def get_schema(server):
    # avoid generating null (\x00) bytes in strings during test case generation
-    return schemathesis.openapi.from_uri(
+    return schemathesis.openapi.from_url(
        f"{server.url_root}/openapi.json",
-        generation_config=GenerationConfig(allow_x00=False),
+        config=SchemathesisConfig(
+            projects=ProjectsConfig(
+                default=ProjectConfig(
+                    generation=GenerationConfig(
+                        allow_x00=False,
+                        modes=[GenerationMode.POSITIVE],
+                    ),
+                    checks=ChecksConfig(
+                        positive_data_acceptance=PositiveDataAcceptanceConfig(
+                            enabled=False,
+                        ),
+                    ),
+                    phases=PhasesConfig(
+                        coverage=CoveragePhaseConfig(enabled=False),
+                    ),
+                ),
+            ),
+        ),
    )


-schema = schemathesis.from_pytest_fixture("get_schema")
+schema = schemathesis.pytest.from_fixture("get_schema")


@schemathesis.hook
-def before_generate_case(context: schemathesis.hooks.HookContext, strategy):
+def before_generate_case(context: schemathesis.HookContext, strategy):
    op = context.operation
    assert op is not None

-    def no_invalid_types(case: schemathesis.models.Case):
+    def no_invalid_types(case: schemathesis.Case):
        """
        Skips tool_calls with `"type": "custom"` which schemathesis incorrectly
        generates instead of the valid `"type": "function"`.
@@ -68,39 +92,25 @@ def before_generate_case(context: schemathesis.hooks.HookContext, strategy):
            -d '{"messages": [{"role": "assistant", "tool_calls": [{"custom": {"input": "", "name": ""}, "id": "", "type": "custom"}]}]}' \
            http://localhost:8000/v1/chat/completions
        """  # noqa: E501
-        if hasattr(case, "body") and isinstance(case.body, dict):
-            if (
-                "messages" in case.body
-                and isinstance(case.body["messages"], list)
-                and len(case.body["messages"]) > 0
-            ):
-                for message in case.body["messages"]:
-                    if not isinstance(message, dict):
-                        continue
+        if (
+            hasattr(case, "body")
+            and isinstance(case.body, dict)
+            and "messages" in case.body
+            and isinstance(case.body["messages"], list)
+            and len(case.body["messages"]) > 0
+        ):
+            for message in case.body["messages"]:
+                if not isinstance(message, dict):
+                    continue

-                    tool_calls = message.get("tool_calls", [])
-                    if isinstance(tool_calls, list):
-                        for tool_call in tool_calls:
-                            if isinstance(tool_call, dict):
-                                if tool_call.get("type") != "function":
-                                    return False
-                                if "custom" in tool_call:
-                                    return False
-
-            # Sometimes structured_outputs.grammar is generated to be empty
-            # Causing a server error in EBNF grammar parsing
-            # https://github.com/vllm-project/vllm/pull/22587#issuecomment-3195253421
-            structured_outputs = case.body.get("structured_outputs", {})
-            grammar = (
-                structured_outputs.get("grammar")
-                if isinstance(structured_outputs, dict)
-                else None
-            )
-
-            if grammar == "":
-                # Allow None (will be handled as no grammar)
-                # But skip empty strings
-                return False
+                tool_calls = message.get("tool_calls", [])
+                if isinstance(tool_calls, list):
+                    for tool_call in tool_calls:
+                        if isinstance(tool_call, dict):
+                            if tool_call.get("type") != "function":
+                                return False
+                            if "custom" in tool_call:
+                                return False

        return True

@@ -108,7 +118,6 @@ def before_generate_case(context: schemathesis.hooks.HookContext, strategy):


@schema.parametrize()
-@schema.override(headers={"Content-Type": "application/json"})
@settings(
    deadline=LONG_TIMEOUT_SECONDS * 1000,
    max_examples=50,
@@ -122,7 +131,7 @@ def before_generate_case(context: schemathesis.hooks.HookContext, strategy):
    # generating large-but-valid request bodies before vLLM is called.
    suppress_health_check=[HealthCheck.filter_too_much, HealthCheck.data_too_large],
 )
-def test_openapi_stateless(case: Case):
+def test_openapi_stateless(case: schemathesis.Case):
    key = (
        case.operation.method.upper(),
        case.operation.path,
@@ -151,4 +160,8 @@ def test_openapi_stateless(case: Case):
    }.get(key, DEFAULT_TIMEOUT_SECONDS)

    # No need to verify SSL certificate for localhost
-    case.call_and_validate(verify=False, timeout=timeout)
+    case.call_and_validate(
+        verify=False,
+        timeout=timeout,
+        headers={"Content-Type": "application/json"},
+    )
@@ -0,0 +1,65 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Tests that validation_exception_handler populates the `param` field
+in its error response using the Pydantic error's `loc`, even when no
+custom VLLMValidationError context is present.
+
+Previously, `param` was only populated for errors carrying a custom
+VLLMValidationError in their Pydantic `ctx`. Plain validation failures
+(missing fields, wrong types) left `param` as None, even though the
+field name was readily available from `error['loc']`.
+"""
+
+import json
+from types import SimpleNamespace
+
+import pytest
+from fastapi.exceptions import RequestValidationError
+
+from vllm.entrypoints.serve.utils.server_utils import validation_exception_handler
+
+
+def _fake_request(log_error_stack: bool = False) -> SimpleNamespace:
+    """Minimal stand-in for a FastAPI Request - just enough for the
+    handler to read req.app.state.args.log_error_stack."""
+    return SimpleNamespace(
+        app=SimpleNamespace(
+            state=SimpleNamespace(args=SimpleNamespace(log_error_stack=log_error_stack))
+        ),
+        state=SimpleNamespace(),  # no request_metadata -> hasattr(...) is False
+    )
+
+
+class TestValidationErrorParamFallback:
+    """Ensure `param` falls back to the Pydantic error's `loc` when no
+    custom VLLMValidationError context is present."""
+
+    @pytest.mark.parametrize(
+        ("error_type", "msg"),
+        [
+            ("missing", "Field required"),
+            ("list_type", "Input should be a valid list"),
+        ],
+        ids=["missing-field", "wrong-type"],
+    )
+    @pytest.mark.asyncio
+    async def test_param_falls_back_to_loc(self, error_type: str, msg: str):
+        errors = [{"type": error_type, "loc": ("body", "messages"), "msg": msg}]
+        exc = RequestValidationError(errors)
+
+        response = await validation_exception_handler(_fake_request(), exc)
+        body = json.loads(response.body)
+
+        assert body["error"]["param"] == "body.messages"
+
+    @pytest.mark.asyncio
+    async def test_param_fallback_does_not_crash_on_non_dict_error(self):
+        """Schemathesis fuzzing found that errors[0] isn't always a dict.
+        The fallback must not crash in that case - it should just leave
+        param as None instead of raising."""
+        exc = RequestValidationError(["some unexpected non-dict error"])
+
+        response = await validation_exception_handler(_fake_request(), exc)
+        body = json.loads(response.body)
+
+        assert body["error"]["param"] is None
@@ -78,7 +78,16 @@ def test_gsm8k_correctness(config_filename):
            "Skipping DeepSeek-V3.2 and DeepSeek-R1 on ROCm platforms "
            "due to agent pool disk space issues and pod evictions."
        )
+    if current_platform.is_rocm() and (
+        "Qwen3.5-35B-A3B-MXFP4-AITER-TP2" in config_filename.name
+    ):
+        from vllm.platforms.rocm import on_gfx950

+        if not on_gfx950():
+            pytest.skip(
+                "Skipping Qwen3.5-35B-A3B-MXFP4-AITER-TP2 on non-GFX950 platforms. "
+                "The quantization scheme is not supported on non-GFX950 platforms."
+            )
    # Parse server arguments from config (use shlex to handle quoted strings)
    server_args_str = eval_config.get("server_args", "")
    server_args = shlex.split(server_args_str) if server_args_str else []
@@ -134,6 +134,7 @@ def _reference_index_topk(
    topk: int,
    init_blocks: int,
    local_blocks: int,
+    sm_scale: float = 1.0,
 ) -> torch.Tensor:
    total_q, num_idx_heads, _ = idx_q.shape
    out = torch.full(
@@ -149,7 +150,7 @@ def _reference_index_topk(
        num_blocks = (seq_len + BLOCK_SIZE - 1) // BLOCK_SIZE
        pages = block_table[req_id, :num_blocks]
        k = index_kv_cache[pages].reshape(num_blocks * BLOCK_SIZE, -1)
-        score = torch.einsum("qhd,kd->hqk", q.float(), k.float())
+        score = torch.einsum("qhd,kd->hqk", q.float(), k.float()) * sm_scale

        q_pos = prefix_len + torch.arange(q_len, device=idx_q.device)
        k_pos = torch.arange(k.shape[0], device=idx_q.device)
@@ -244,6 +245,270 @@ def test_prefill_index_topk_correctness():
    _assert_topk_indices_equal_unordered(actual, expected)


+# MSA indexer (SM100): fmha_sm100 OnlyScore for the per-block scores, then the
+# Triton minimax_m3_index_topk for selection (no sparse_topk_select). Uses a
+# deterministic construction (idx_q == 1, distinct e4m3-exact per-block values)
+# so scores are strictly monotonic in the block id -> exact top-k agreement.
+def _fmha_indexer_topk(
+    idx_q: torch.Tensor,  # [total_q, H, 128] bf16/e4m3
+    index_cache: torch.Tensor,  # [num_pages, 128, 128] bf16/e4m3
+    block_table: torch.Tensor,
+    q_lens: torch.Tensor,
+    seq_lens: torch.Tensor,
+    prefix_lens: torch.Tensor,
+    sm_scale: float,
+    topk: int,
+) -> torch.Tensor:
+    """Replicate MiniMaxM3IndexerMSAImpl's score path (single decode/prefill side)."""
+    from vllm.third_party.fmha_sm100.api import _fmha_sm100, _fmha_sm100_plan
+
+    num_idx_heads, head_dim = idx_q.shape[1], idx_q.shape[2]
+    nvp = [(s + 127) // 128 for s in seq_lens.tolist()]
+    kv_indices = torch.cat([block_table[r, : nvp[r]] for r in range(len(nvp))]).to(
+        torch.int32
+    )
+
+    qo = q_lens.cpu().to(torch.int32)
+    kv = seq_lens.cpu().to(torch.int32)
+    plan = _fmha_sm100_plan(
+        qo,
+        kv,
+        num_idx_heads,
+        num_kv_heads=1,
+        qo_offset=kv - qo,
+        page_size=128,
+        output_maxscore=True,
+        causal=True,
+        num_kv_splits=1,
+    )
+    k_pages = index_cache.view(index_cache.shape[0], 1, 128, head_dim)
+    _, max_score = _fmha_sm100(
+        idx_q,
+        k_pages,
+        k_pages,
+        plan,
+        kv_indices=kv_indices,
+        output_o=False,
+        output_maxscore=True,
+        sm_scale=sm_scale,
+    )
+
+    batch = q_lens.numel()
+    cu = torch.zeros(batch + 1, dtype=torch.int32, device=idx_q.device)
+    cu[1:] = q_lens.to(torch.int32).cumsum(0)
+    # max_score [H, k_tiles, total_q] -> transpose to [H, total_q, k_tiles].
+    return minimax_m3_index_topk(
+        max_score.transpose(1, 2),
+        cu,
+        prefix_lens.to(torch.int32),
+        int(q_lens.max()),
+        topk,
+        0,  # init_blocks
+        0,  # local_blocks
+    )
+
+
+# e4m3-exact, strictly-increasing per-block values: with idx_q == 1 (also exact)
+# the per-block scores are exact and distinct in BOTH bf16 and e4m3, so the fp8
+# score path selects the same top-k as the reference (no quantization ties).
+_E4M3_EXACT_VALUES = [
+    *range(1, 17),  # 1..16  (step 1)
+    *range(18, 33, 2),  # 18..32 (step 2)
+    *range(36, 65, 4),  # 36..64 (step 4)
+    *range(72, 129, 8),  # 72..128 (step 8)
+]
+
+
+@pytest.mark.skipif(
+    not current_platform.is_device_capability_family(100),
+    reason="fmha_sm100 indexer requires SM100 (Blackwell).",
+)
+@pytest.mark.parametrize("index_dtype", [torch.bfloat16, torch.float8_e4m3fn])
+@pytest.mark.parametrize(
+    ("q_lens", "prefix_lens"),
+    [
+        ((4, 3), (2048, 2560)),  # prefill: every token sees >= 16 causal blocks
+        ((1, 1, 1), (2048, 3000, 4096)),  # decode: one query token per request
+    ],
+)
+def test_fmha_sm100_indexer_matches_reference(q_lens, prefix_lens, index_dtype):
+    torch.manual_seed(0)
+    num_idx_heads, head_dim = 4, HEAD_DIM
+    device = "cuda"
+
+    q_lens_t = torch.tensor(q_lens, device=device, dtype=torch.int32)
+    prefix_lens_t = torch.tensor(prefix_lens, device=device, dtype=torch.int32)
+    seq_lens = prefix_lens_t + q_lens_t
+    batch = len(q_lens)
+    max_blocks = (int(seq_lens.max()) + BLOCK_SIZE - 1) // BLOCK_SIZE
+    assert max_blocks <= len(_E4M3_EXACT_VALUES)
+    num_pages = batch * max_blocks
+    block_table = torch.randperm(num_pages, device=device, dtype=torch.int32).reshape(
+        batch, max_blocks
+    )
+
+    idx_q = torch.ones(
+        int(q_lens_t.sum()), num_idx_heads, head_dim, device=device, dtype=index_dtype
+    )
+    index_cache = torch.empty(
+        num_pages, BLOCK_SIZE, head_dim, device=device, dtype=index_dtype
+    )
+    for r in range(batch):
+        for b in range(max_blocks):
+            index_cache[block_table[r, b]] = float(_E4M3_EXACT_VALUES[b])
+
+    sm_scale = head_dim**-0.5
+    actual = _fmha_indexer_topk(
+        idx_q,
+        index_cache,
+        block_table,
+        q_lens_t,
+        seq_lens,
+        prefix_lens_t,
+        sm_scale,
+        TOPK,
+    )
+    expected = _reference_index_topk(
+        idx_q,
+        index_cache,
+        block_table,
+        q_lens_t,
+        seq_lens,
+        prefix_lens_t,
+        TOPK,
+        init_blocks=0,
+        local_blocks=0,
+        sm_scale=sm_scale,
+    )
+    _assert_topk_indices_equal_unordered(actual, expected)
+
+
+# Full impl-level parity: drive both MiniMaxM3IndexerMSAImpl (fmha_sm100 score +
+# Triton top-k) and MiniMaxM3IndexerTritonImpl through their real metadata
+# builders on the SAME CommonAttentionMetadata + index cache, and assert the
+# selected blocks agree. This exercises all the metadata the impl/kernels consume
+# (decode/prefill split, cu_seqlens_q rebasing, prefix_lens, kv_indices gather,
+# decode_pages split) -- a metadata bug on either side shifts the causal window
+# or the block->page mapping and breaks the comparison.
+@pytest.mark.skipif(
+    not current_platform.is_device_capability_family(100),
+    reason="fmha_sm100 indexer requires SM100 (Blackwell).",
+)
+@pytest.mark.parametrize("topk", [8, 16])
+def test_msa_indexer_impl_matches_triton(topk, monkeypatch):
+    import vllm.models.minimax_m3.common.indexer as indexer_mod
+    from tests.v1.attention.utils import (
+        BatchSpec,
+        create_common_attn_metadata,
+        create_vllm_config,
+    )
+    from vllm.config import set_current_vllm_config
+    from vllm.forward_context import set_forward_context
+    from vllm.models.minimax_m3.common.indexer import (
+        MiniMaxM3IndexerTritonImpl,
+        MiniMaxM3IndexerTritonMetadataBuilder,
+    )
+    from vllm.models.minimax_m3.nvidia.indexer_msa import (
+        MiniMaxM3IndexerMSAImpl,
+        MiniMaxM3IndexerMSAMetadataBuilder,
+    )
+
+    torch.manual_seed(0)
+    device = torch.device("cuda")
+    num_idx_heads, head_dim = 4, HEAD_DIM
+    # TP=1: avoid requiring an initialized distributed group in a unit test.
+    monkeypatch.setattr(indexer_mod, "get_tensor_model_parallel_world_size", lambda: 1)
+
+    vllm_config = create_vllm_config(
+        block_size=BLOCK_SIZE, max_model_len=8192, max_num_batched_tokens=8192
+    )
+    vllm_config.model_config.hf_config.sparse_attention_config = {
+        "sparse_num_index_heads": num_idx_heads
+    }
+
+    # Decode-first mixed batch: 2 decode reqs (q_len 1) then 2 prefill reqs. Long
+    # prefixes so every token sees > TOPK causal blocks (non-trivial selection).
+    batch = BatchSpec(seq_lens=[2305, 2561, 2624, 2720], query_lens=[1, 1, 64, 96])
+    common = create_common_attn_metadata(
+        batch, BLOCK_SIZE, device, arange_block_indices=True
+    )
+    num_tokens = batch.compute_num_tokens()
+
+    # Deterministic index cache: distinct, monotonic per-logical-block values so
+    # the top-k is unambiguous (both kernels pick the same blocks, no fp ties).
+    block_table = common.block_table_tensor
+    num_pages = int(block_table.max().item()) + 1
+    index_cache = torch.zeros(
+        num_pages, BLOCK_SIZE, head_dim, device=device, dtype=DTYPE
+    )
+    for r, seq_len in enumerate(batch.seq_lens):
+        for b in range((seq_len + BLOCK_SIZE - 1) // BLOCK_SIZE):
+            index_cache[block_table[r, b]] = float(b + 1)
+    index_q = torch.ones(
+        num_tokens, num_idx_heads * head_dim, device=device, dtype=DTYPE
+    )
+
+    spec = MLAAttentionSpec(
+        block_size=BLOCK_SIZE, num_kv_heads=1, head_size=head_dim, dtype=DTYPE
+    )
+    impl_kwargs = dict(
+        num_kv_heads=num_idx_heads,
+        scale=head_dim**-0.5,
+        topk_blocks=topk,
+        sparse_block_size=BLOCK_SIZE,
+        num_index_heads=num_idx_heads,
+        index_head_dim=head_dim,
+        init_blocks=0,
+        local_blocks=0,
+    )
+
+    with set_current_vllm_config(vllm_config):
+        msa_impl = MiniMaxM3IndexerMSAImpl(prefix="idx_msa", **impl_kwargs)
+        triton_impl = MiniMaxM3IndexerTritonImpl(prefix="idx_triton", **impl_kwargs)
+        msa_builder = MiniMaxM3IndexerMSAMetadataBuilder(
+            spec, [msa_impl.index_cache.prefix], vllm_config, device
+        )
+        triton_builder = MiniMaxM3IndexerTritonMetadataBuilder(
+            spec, [triton_impl.index_cache.prefix], vllm_config, device
+        )
+
+    # Both impls score against the same index keys.
+    msa_impl.index_cache.kv_cache = index_cache
+    triton_impl.index_cache.kv_cache = index_cache
+
+    # Exercise the shared persistent top-k buffer for BOTH impls: each must write
+    # decode ([:, :nd]) and prefill ([:, nd:]) into its buffer and return views.
+    # Separate buffers so the two forwards don't clobber each other.
+    nd = sum(q for q in batch.query_lens if q <= 1)
+    msa_impl.topk_indices_buffer = torch.full(
+        (num_idx_heads, num_tokens, topk), -2, dtype=torch.int32, device=device
+    )
+    triton_impl.topk_indices_buffer = torch.full(
+        (num_idx_heads, num_tokens, topk), -2, dtype=torch.int32, device=device
+    )
+
+    attn_metadata = {
+        msa_impl.index_cache.prefix: msa_builder.build(0, common),
+        triton_impl.index_cache.prefix: triton_builder.build(0, common),
+    }
+    with set_forward_context(attn_metadata, vllm_config):
+        msa_decode, msa_prefill = msa_impl(index_q)
+        tri_decode, tri_prefill = triton_impl(index_q)
+
+    assert msa_decode is not None and tri_decode is not None
+    assert msa_prefill is not None and tri_prefill is not None
+    _assert_topk_indices_equal_unordered(msa_decode, tri_decode)
+    _assert_topk_indices_equal_unordered(msa_prefill, tri_prefill)
+    # decode/prefill outputs are views into each impl's persistent buffer.
+    for impl, dec, pre in (
+        (msa_impl, msa_decode, msa_prefill),
+        (triton_impl, tri_decode, tri_prefill),
+    ):
+        buf = impl.topk_indices_buffer
+        assert dec.data_ptr() == buf[:, :nd, :].data_ptr()
+        assert pre.data_ptr() == buf[:, nd:, :].data_ptr()
+
+
@pytest.mark.parametrize(
    ("decode_query_len", "max_decode_query_len"),
    [
@@ -317,6 +582,65 @@ def test_decode_index_topk_correctness(
    _assert_topk_indices_equal_unordered(actual, expected)


+@pytest.mark.skipif(
+    not current_platform.is_device_capability_family(100),
+    reason="fp8 e4m3 indexer cache is the SM100 (MSA) path.",
+)
+@pytest.mark.parametrize("num_idx_heads", [1, 4])
+def test_decode_index_topk_fp8(num_idx_heads: int):
+    """The fp8 (e4m3) indexer cache feeds the Triton decode kernel on the MSA
+    path. The kernel must score in fp32 (no scaling) so its top-k matches a
+    reference computed from the dequantized fp8 values."""
+    torch.manual_seed(0)
+    topk, init_blocks, local_blocks, head_dim = 8, 0, 1, 128
+    decode_query_len = 1
+    active_seq_lens = torch.tensor((129, 1025, 4097), device="cuda", dtype=torch.int32)
+    q_lens = torch.full_like(active_seq_lens, decode_query_len)
+    prefix_lens = active_seq_lens - decode_query_len
+    batch = active_seq_lens.numel()
+    max_seq_len = int(active_seq_lens.max())
+    max_blocks = (max_seq_len + BLOCK_SIZE - 1) // BLOCK_SIZE
+    num_pages = batch * max_blocks
+    block_table = torch.randperm(num_pages, device="cuda", dtype=torch.int32).reshape(
+        batch, max_blocks
+    )
+    idx_q = torch.randn(
+        batch * decode_query_len, num_idx_heads, head_dim, device="cuda"
+    ).to(torch.float8_e4m3fn)
+    index_kv_cache = torch.randn(num_pages, BLOCK_SIZE, head_dim, device="cuda").to(
+        torch.float8_e4m3fn
+    )
+
+    actual = minimax_m3_index_decode(
+        idx_q,
+        index_kv_cache,
+        block_table,
+        active_seq_lens,
+        max_seq_len=max_seq_len,
+        topk=topk,
+        init_blocks=init_blocks,
+        local_blocks=local_blocks,
+        num_kv_heads=num_idx_heads,
+        sm_scale=head_dim**-0.5,
+        decode_query_len=decode_query_len,
+    )
+    # Reference from the DEQUANTIZED fp8 values (the kernel computes the fp8 QK
+    # in fp32, so it must match an fp32 matmul of the same e4m3 values).
+    expected = _reference_index_topk(
+        idx_q.float(),
+        index_kv_cache.float(),
+        block_table,
+        q_lens,
+        active_seq_lens,
+        prefix_lens,
+        topk,
+        init_blocks,
+        local_blocks,
+        head_dim**-0.5,
+    )
+    _assert_topk_indices_equal_unordered(actual, expected)
+
+
 # Sparse attention kernels.
 def _reference_sparse_attn(
    q: torch.Tensor,
@@ -5,8 +5,9 @@ Tests for the FlashInfer TRTLLM NvFP4 MoE backend
 (`TrtLlmNvFp4ExpertsModular`).

 Covers the activations the wrapper claims to support — SiLU, RELU^2 (non-gated),
-and GELU — including a Gemma4-shaped case (128 experts, top-k 8,
-intermediate_size 704) that exercises the non-256-aligned padding path.
+GELU, and clamped SwiGLU-OAI (MiniMax-M3) — including a Gemma4-shaped case
+(128 experts, top-k 8, intermediate_size 704) that exercises the non-256-aligned
+padding path.
 """

 import pytest
@@ -80,6 +81,29 @@ if _CLAMP_OP_NAME not in op_registry:

 SILU_WITH_CLAMP = op_registry[_CLAMP_OP_NAME]

+# Clamped SwiGLU-OAI (MiniMax-M3): non-default alpha/beta so the kernel must
+# honor gemm1_alpha (raw) and gemm1_beta (folded by g1_alphas), not just clamp.
+_SWIGLU_ALPHA = 1.702
+_SWIGLU_BETA = 1.0
+_OAI_OP_NAME = "test_swigluoai_with_clamp"
+
+if _OAI_OP_NAME not in op_registry:
+
+    @CustomOp.register(_OAI_OP_NAME)
+    class _SwigluOAIWithClampTest(SiluAndMulWithClamp):
+        custom_op_name = _OAI_OP_NAME
+
+        def __init__(self, *, compile_native: bool = True) -> None:
+            super().__init__(
+                _SWIGLU_LIMIT,
+                alpha=_SWIGLU_ALPHA,
+                beta=_SWIGLU_BETA,
+                compile_native=compile_native,
+            )
+
+
+SWIGLUOAI_REF = op_registry[_OAI_OP_NAME]
+

 ACTIVATION_CASES = [
    pytest.param(MoEActivation.SILU, MoEActivation.SILU, None, id="silu"),
@@ -91,6 +115,12 @@ ACTIVATION_CASES = [
        id="relu2_no_mul",
    ),
    pytest.param(MoEActivation.GELU, MoEActivation.GELU, None, id="gelu"),
+    pytest.param(
+        MoEActivation.SWIGLUOAI_UNINTERLEAVE,
+        SWIGLUOAI_REF,
+        _SWIGLU_LIMIT,
+        id="swigluoai_uninterleave",
+    ),
 ]


@@ -148,6 +178,10 @@ def test_trtllm_fp4_moe_no_graph(
            is_scale_swizzled=False,
        )
        quant_config.gemm1_clamp_limit = swiglu_limit
+        is_oai = activation == MoEActivation.SWIGLUOAI_UNINTERLEAVE
+        if is_oai:
+            quant_config.gemm1_alpha = _SWIGLU_ALPHA
+            quant_config.gemm1_beta = _SWIGLU_BETA
        if swiglu_limit is not None:
            assert quant_config.g1_alphas is not None
            assert quant_config.a2_gscale is not None
@@ -192,6 +226,27 @@ def test_trtllm_fp4_moe_no_graph(
        fake_layer.w2_input_scale = torch.ones_like(quant_config.g2_alphas)
        trtllm_inner.process_weights_after_loading(fake_layer)

+        if is_oai:
+            # alpha stays raw; beta and clamp are folded by g1_alphas
+            # (== _LARGE_OUTPUT1_SCALE here), so the fold is load-bearing.
+            assert torch.allclose(
+                trtllm_inner.gemm1_alpha,
+                torch.full_like(trtllm_inner.gemm1_alpha, _SWIGLU_ALPHA),
+            )
+            assert torch.allclose(
+                trtllm_inner.gemm1_beta,
+                torch.full_like(
+                    trtllm_inner.gemm1_beta, _SWIGLU_BETA / _LARGE_OUTPUT1_SCALE
+                ),
+            )
+            assert torch.allclose(
+                trtllm_inner.gemm1_clamp_limit,
+                torch.full_like(
+                    trtllm_inner.gemm1_clamp_limit,
+                    _SWIGLU_LIMIT / _LARGE_OUTPUT1_SCALE,
+                ),
+            )
+
        trtllm_experts = mk.FusedMoEKernel(
            maybe_make_prepare_finalize(
                moe=moe_config,
@@ -345,6 +345,63 @@ def test_per_token_group_quant_fp8_packed_zero_fills_padded_output_q(
        )


+@pytest.mark.skipif(
+    not current_platform.is_cuda_alike(),
+    reason="packed FP8 per-token-group quant kernel requires a CUDA-alike GPU",
+)
+def test_per_token_group_quant_fp8_packed_large_mn():
+    """Regression test for https://github.com/vllm-project/vllm/issues/45099.
+
+    Some background: gridDim.x and gridDim.y have different limits of 2^31 - 1 and
+    2^16 - 1, respectively.
+    Prior code introduced a bug where it incorrectly assumed grid.x and y both have
+    2^31 - 1 limits and mixed them up, which doesn't surface until the kernel is
+    launched with a large mn that exceeds grid.y limit (2^16 - 1).
+
+    This issue doesn't surface often because each forward pass only processes a
+    bounded token batch, not the full context.
+    Quantizing tensors with more rows than that will fail at launch with
+    "CUDA error: invalid argument".
+    This is a differential test that compares fp8 output against Triton output
+    reference when token size sits just above the gridDim.y 2^16 - 1 limit.
+    """
+
+    device = "cuda"
+    group_size = 128
+    # hidden 2048 -> 2048/128 = 16 groups per row -> kx=16, ry=1: one grid row per mn
+    # row, so any mn > 65535 overflowed grid.y before the fix.
+    num_tokens, hidden_dim = 65537, 2048
+    torch.manual_seed(42)
+    x = torch.randn((num_tokens, hidden_dim), device=device, dtype=torch.bfloat16) * 8
+
+    out_q, out_s_packed = fp8_utils.per_token_group_quant_fp8_packed_for_deepgemm(
+        x,
+        group_size=group_size,
+        use_ue8m0=True,
+    )
+
+    with patch("vllm.platforms.current_platform.is_cuda_alike", return_value=False):
+        ref_q, ref_s = fp8_utils.per_token_group_quant_fp8(
+            x, group_size, use_ue8m0=True
+        )
+
+    assert torch.equal(out_q, ref_q), "Quantized output mismatch"
+
+    # Vectorized packed-scale check; the per-element loop used by the smaller
+    # tests is too slow at this size. groups_per_row is a multiple of 4 here,
+    # so there is no K padding and the packed view lines up.
+    mn = num_tokens
+    groups_per_row = hidden_dim // group_size
+    k_num_packed = (groups_per_row + 3) // 4
+    assert groups_per_row % 4 == 0
+    ref_exponents = (ref_s.reshape(mn, groups_per_row).view(torch.int32) >> 23) & 0xFF
+    exp = ref_exponents.view(mn, k_num_packed, 4)
+    expected = (
+        exp[..., 0] | (exp[..., 1] << 8) | (exp[..., 2] << 16) | (exp[..., 3] << 24)
+    )
+    assert torch.equal(out_s_packed.cpu(), expected.cpu()), "Packed scale mismatch"
+
+
@pytest.mark.parametrize("shape", [(32, 128), (64, 256), (16, 512)])
@pytest.mark.parametrize("group_size", [64, 128])
@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA not available")
@@ -60,8 +60,10 @@ def test_rocm_compressed_tensors_w8a8(
    vllm_runner, example_prompts, model_path, max_tokens, num_logprobs
 ):
    dtype = "bfloat16"
-
-    with vllm_runner(model_path, dtype=dtype) as vllm_model:
+    # Pin to TRITON_ATTN, see https://github.com/vllm-project/vllm/issues/46179
+    with vllm_runner(
+        model_path, dtype=dtype, attention_backend="TRITON_ATTN"
+    ) as vllm_model:
        vllm_model.generate_greedy_logprobs(example_prompts, max_tokens, num_logprobs)


@@ -278,3 +278,99 @@ def test_sparse_full(num_tokens, block_size, kv_cache_dtype):
    torch.testing.assert_close(
        index_cache.view(-1, HEAD_DIM), expected_index_cache, rtol=0, atol=0
    )
+
+
+# ── Test 3: fp8 (e4m3) index outputs ─────────────────────────────────────────
+# The fp8 score path stores index_q and the index-K cache as e4m3 while q/k/v +
+# q_out stay bf16. Asserts: (1) q/k/v/q_out are bit-identical to the bf16 run
+# (the index dtype must not perturb the main branch), and (2) the e4m3 index
+# outputs dequantize close to the bf16 reference.
+
+
+@pytest.mark.skipif(
+    not torch.cuda.is_available() or torch.cuda.get_device_capability() < (8, 9),
+    reason="e4m3 conversion requires CUDA SM89+.",
+)
+@pytest.mark.parametrize("num_tokens", [1, 7, 64, 513])
+@pytest.mark.parametrize("block_size", [16, 64])
+def test_sparse_full_fp8_index(num_tokens, block_size):
+    torch.manual_seed(1)
+    device, dtype, eps = "cuda", torch.bfloat16, 1e-6
+    base, max_pos = 5_000_000.0, 4096
+    num_heads, num_kv_heads, num_idx_heads = 16, 4, 4
+
+    q_w = torch.randn(HEAD_DIM, dtype=dtype, device=device) * 0.1
+    k_w = torch.randn(HEAD_DIM, dtype=dtype, device=device) * 0.1
+    iq_w = torch.randn(HEAD_DIM, dtype=dtype, device=device) * 0.1
+    ik_w = torch.randn(HEAD_DIM, dtype=dtype, device=device) * 0.1
+    cos_sin = make_cos_sin_cache(max_pos, ROTARY_DIM, base, dtype, device)
+    positions = torch.randint(
+        0, max_pos, (num_tokens,), dtype=torch.int64, device=device
+    )
+
+    qsz, kvsz = num_heads * HEAD_DIM, num_kv_heads * HEAD_DIM
+    iqsz, iksz = num_idx_heads * HEAD_DIM, HEAD_DIM
+    qkv0 = torch.randn(
+        num_tokens, qsz + 2 * kvsz + iqsz + iksz, dtype=dtype, device=device
+    )
+
+    num_blocks = (num_tokens + block_size - 1) // block_size + 1
+    slot_mapping = torch.randperm(
+        num_blocks * block_size, dtype=torch.int64, device=device
+    )[:num_tokens]
+    index_slot_mapping = torch.roll(slot_mapping, shifts=1)
+
+    def run(index_dtype):
+        qkv = qkv0.clone()
+        kv_cache = torch.zeros(
+            num_blocks,
+            2,
+            block_size,
+            num_kv_heads,
+            HEAD_DIM,
+            dtype=dtype,
+            device=device,
+        )
+        index_cache = torch.zeros(
+            num_blocks, block_size, HEAD_DIM, dtype=index_dtype, device=device
+        )
+        q_out = torch.empty(num_tokens, qsz, dtype=dtype, device=device)
+        index_q = torch.empty(num_tokens, iqsz, dtype=index_dtype, device=device)
+        ops.fused_minimax_m3_qknorm_rope_kv_insert(
+            qkv,
+            q_w,
+            k_w,
+            cos_sin,
+            positions,
+            num_heads,
+            num_kv_heads,
+            ROTARY_DIM,
+            eps,
+            iq_w,
+            ik_w,
+            num_idx_heads,
+            slot_mapping,
+            index_slot_mapping,
+            kv_cache,
+            index_cache,
+            block_size,
+            q_out,
+            index_q,
+        )
+        return qkv, kv_cache, index_cache, q_out, index_q
+
+    qkv_bf, kvc_bf, idxc_bf, qo_bf, iq_bf = run(torch.bfloat16)
+    qkv_fp, kvc_fp, idxc_fp, qo_fp, iq_fp = run(torch.float8_e4m3fn)
+
+    assert iq_fp.dtype == torch.float8_e4m3fn
+    assert idxc_fp.dtype == torch.float8_e4m3fn
+
+    # (1) The main branch (q/k/v in qkv, q_out, kv cache) must be bit-identical:
+    # the index output dtype must not perturb anything else.
+    torch.testing.assert_close(qo_fp, qo_bf, rtol=0, atol=0)
+    torch.testing.assert_close(qkv_fp, qkv_bf, rtol=0, atol=0)
+    torch.testing.assert_close(kvc_fp, kvc_bf, rtol=0, atol=0)
+
+    # (2) Dequantized e4m3 index outputs match the bf16 reference within fp8 ulp.
+    torch.testing.assert_close(iq_fp.float(), iq_bf.float(), rtol=0.13, atol=0.05)
+    torch.testing.assert_close(idxc_fp.float(), idxc_bf.float(), rtol=0.13, atol=0.05)
@@ -482,3 +482,127 @@ def test_kernels_hidden_size(
            seq_length=128,
            add_inputs=True,
        )
+
+
+@pytest.mark.parametrize("device", DEVICES)
+def test_add_lora_fused_moe_early_exit(device):
+    """
+    Ensures add_lora_fused_moe does not invoke the LoRA kernel or
+    modify the output tensor when no_lora_flag_cpu is True
+    """
+    from types import SimpleNamespace
+
+    from vllm.lora.punica_wrapper.punica_gpu import PunicaWrapperGPU
+
+    torch.set_default_device(device)
+    torch.accelerator.set_device_index(device)
+
+    max_loras, num_tokens = 4, 16
+    num_experts, top_k, max_lora_rank = 8, 2, 16
+    K, N = 256, 128
+
+    # build PunicaWrapperGPU with minimal lora_config mock
+    lora_config = SimpleNamespace(
+        max_loras=max_loras,
+        specialize_active_lora=False,
+    )
+    wrapper = PunicaWrapperGPU(
+        max_num_batched_tokens=num_tokens,
+        max_batches=num_tokens,
+        device=device,
+        lora_config=lora_config,
+    )
+
+    # simulate a prior LoRA batch so the internal mapping is
+    # populated with stale LoRA IDs
+    lora_mapping = torch.zeros(
+        num_tokens,
+        dtype=torch.int32,
+        device=device,
+    )
+    lora_mapping[:8] = 1
+    lora_mapping[8:] = 2
+    wrapper.token_mapping_meta.prepare_tensors(lora_mapping)
+
+    # simulate a base-model batch (all -1)
+    base_mapping = torch.full(
+        (num_tokens,),
+        -1,
+        dtype=torch.int32,
+        device=device,
+    )
+    wrapper.token_mapping_meta.prepare_tensors(base_mapping)
+
+    assert wrapper.token_mapping_meta.no_lora_flag_cpu[0].item() is True
+
+    # dummy tensors for add_lora_fused_moe
+    y = torch.rand(num_tokens, top_k, N, dtype=torch.bfloat16, device=device)
+    y_snapshot = y.clone()
+    x = torch.rand(num_tokens, K, dtype=torch.bfloat16, device=device)
+
+    lora_a_stacked = (
+        torch.rand(
+            max_loras,
+            num_experts,
+            max_lora_rank,
+            K,
+            dtype=torch.bfloat16,
+            device=device,
+        ),
+    )
+    lora_b_stacked = (
+        torch.rand(
+            max_loras,
+            num_experts,
+            N,
+            max_lora_rank,
+            dtype=torch.bfloat16,
+            device=device,
+        ),
+    )
+    topk_weights = torch.ones(
+        num_tokens,
+        top_k,
+        dtype=torch.float32,
+        device=device,
+    )
+    adapter_enabled = torch.ones(
+        max_loras + 1,
+        dtype=torch.int32,
+        device=device,
+    )
+    shrink_config = expand_config = {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "NUM_WARPS": 4,
+        "NUM_STAGES": 3,
+        "SPLIT_K": 1,
+    }
+
+    # call add_lora_fused_moe - the early exit should prevent any
+    # modification to the output
+    wrapper.add_lora_fused_moe(
+        y=y,
+        x=x,
+        lora_a_stacked=lora_a_stacked,
+        lora_b_stacked=lora_b_stacked,
+        topk_weights=topk_weights,
+        sorted_token_ids=None,
+        expert_ids=torch.zeros(
+            num_tokens * top_k,
+            dtype=torch.int32,
+            device=device,
+        ),
+        num_tokens_post_padded=None,
+        max_lora_rank=max_lora_rank,
+        top_k_num=top_k,
+        shrink_config=shrink_config,
+        expand_config=expand_config,
+        adapter_enabled=adapter_enabled,
+    )
+
+    assert torch.equal(y, y_snapshot), (
+        "add_lora_fused_moe modified output tensor despite no_lora_flag_cpu=True"
+    )
@@ -130,8 +130,12 @@ def test_models(
        monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
        if model == "TitanML/tiny-mixtral":
            # Untrained model: near-uniform logits make argmax sensitive to
-            # AITER's bfloat16 rounding error in plain rms_norm.
+            # AITER's bfloat16 rounding error. Route the plain rms_norm and the
+            # fused MoE (whose near-uniform router logits flip expert selection
+            # under ~1 ULP drift) through the native kernels for this model.
+            # See ROCm/aiter#3806 for the tracking issue and minimal repro.
            monkeypatch.setenv("VLLM_ROCM_USE_AITER_RMSNORM", "0")
+            monkeypatch.setenv("VLLM_ROCM_USE_AITER_MOE", "0")
    elif use_rocm_aiter and model not in AITER_MODEL_LIST:
        # Skip model that are not using AITER tests.
        # When more AITER kernels are added, this list will not be
@@ -25,7 +25,7 @@ TEST_IMAGE_NAMES = [
 ]
 MAX_MODEL_LEN = 8192
 REQUESTS_PER_ROUND = 4
-WARMUP_ROUNDS = 1
+WARMUP_ROUNDS = 2
 MEASURED_ROUNDS = 16
 GPU_GROWTH_THRESHOLD_MIB = 0
 CPU_PEAK_GROWTH_THRESHOLD_MIB = 0
@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 import pytest
+import torch
 import torch.nn.functional as F
 import transformers.utils
 from PIL import Image
@@ -52,6 +53,7 @@ def _get_cherry_blossom_image() -> Image.Image:
    )


+@torch.inference_mode()
 def _run_test(
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
@@ -92,3 +92,49 @@ def test_processor_num_frames_timestamp(
    assert len(video_phs) == 1, (
        f"Expected exactly 1 video placeholder, got {len(video_phs)}"
    )
+
+
+@pytest.mark.parametrize("model_id", [MODEL_ID])
+@pytest.mark.parametrize("num_videos", [2, 4])
+def test_processor_multi_video(
+    model_id: str,
+    num_videos: int,
+) -> None:
+    """Verify that multi-video processing produces correct placeholders.
+
+    This exercises the token-level replacement path in
+    ``_call_hf_processor`` which avoids the quadratic text-level
+    prompt expansion.
+    """
+    ctx = build_model_context(
+        model_id,
+        limit_mm_per_prompt={"image": 0, "video": num_videos},
+    )
+    processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config)
+
+    prompt = "<|vision_start|><|video_pad|><|vision_end|>" * num_videos
+    mm_data = {"video": [_build_video_mm_data(num_frames=8)["video"][0]] * num_videos}
+
+    processed = processor(
+        prompt,
+        mm_items=processor.info.parse_mm_data(mm_data),
+        hf_processor_mm_kwargs={"num_frames": 8},
+    )
+
+    token_ids = processed["prompt_token_ids"]
+    assert len(token_ids) > 0
+
+    video_phs = processed["mm_placeholders"].get("video", [])
+    assert len(video_phs) == num_videos, (
+        f"Expected {num_videos} video placeholders, got {len(video_phs)}"
+    )
+
+    # All placeholders should have the same length (same video params)
+    # and must not overlap.
+    lengths = {ph.length for ph in video_phs}
+    assert len(lengths) == 1, f"Placeholder lengths differ: {lengths}"
+    for i in range(1, len(video_phs)):
+        prev_end = video_phs[i - 1].offset + video_phs[i - 1].length
+        assert video_phs[i].offset >= prev_end, (
+            f"Placeholder {i} overlaps with placeholder {i - 1}"
+        )
@@ -96,6 +96,9 @@ class MockTokenizer:
        return "".join(parts)


+CHUNK_SIZES = [1, 2, 3, 5, 11, 23, None]
+
+
 def make_mock_tokenizer(sample: Sample) -> MockTokenizer:
    """Build a mock tokenizer from a sample's vocab and token data."""
    return MockTokenizer(
@@ -19,6 +19,7 @@ import pytest
 from pydantic import TypeAdapter

 from tests.parser.engine.replay_harness import (
+    CHUNK_SIZES,
    MockTokenizer,
    assert_parse_output,
    collect_output,
@@ -113,8 +114,6 @@ _PAIRINGS = _discover_pairings()

 _ALL_SAMPLES = [(p.parser_cls, s) for p in _PAIRINGS for s in p.samples]

-CHUNK_SIZES = [1, 2, 3, 5, 11, 23, None]
-

@pytest.mark.parametrize("chunk_size", CHUNK_SIZES, ids=lambda c: f"chunk={c}")
@pytest.mark.parametrize(
@@ -0,0 +1,181 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Regression test for U+FFFD leak at reasoning→content transition.
+
+When byte-fallback tokens span the reasoning/content boundary,
+decoding isolated content-side token IDs via tokenizer.decode()
+produces U+FFFD (Unicode replacement character).  The fix flushes
+the reasoning parser's engine lexer instead.
+
+Reproduces the bug at various chunk sizes and validates that the
+fix prevents U+FFFD from leaking into streamed content.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from tests.parser.engine.replay_harness import (
+    CHUNK_SIZES,
+    MockTokenizer,
+    collect_output,
+    replay_streaming,
+)
+from vllm.parser.abstract_parser import DelegatingParser
+from vllm.parser.engine.registered_adapters import (
+    Glm47MoeParserReasoningAdapter,
+    Glm47MoeParserToolAdapter,
+    Qwen3ParserReasoningAdapter,
+    Qwen3ParserToolAdapter,
+)
+
+
+class ByteFallbackMockTokenizer(MockTokenizer):
+    """MockTokenizer that returns U+FFFD for specified token IDs.
+
+    Simulates byte-fallback tokenizer behavior where isolated
+    partial-byte tokens decode to the Unicode replacement character.
+    """
+
+    def __init__(
+        self,
+        vocab: dict[str, int],
+        tokens: list[tuple[int, str]],
+        ufffd_token_ids: set[int],
+    ) -> None:
+        super().__init__(vocab, tokens)
+        self._ufffd_token_ids = frozenset(ufffd_token_ids)
+
+    def decode(self, ids: list[int], skip_special_tokens: bool = False) -> str:
+        parts: list[str] = []
+        for tid in ids:
+            if skip_special_tokens and tid in self._special_ids:
+                continue
+            if tid in self._ufffd_token_ids:
+                parts.append("�")
+            else:
+                text = self._token_decode_map.get(tid, f"?{tid}?")
+                parts.append(text)
+        return "".join(parts)
+
+
+# ── Model-specific DelegatingParser subclasses ───────────────────────
+
+
+class _Glm47Delegating(DelegatingParser):
+    reasoning_parser_cls = Glm47MoeParserReasoningAdapter
+    tool_parser_cls = Glm47MoeParserToolAdapter
+
+
+class _Qwen3Delegating(DelegatingParser):
+    reasoning_parser_cls = Qwen3ParserReasoningAdapter
+    tool_parser_cls = Qwen3ParserToolAdapter
+
+
+# ── Shared test data ─────────────────────────────────────────────────
+
+_SHARED_TOKENS: list[tuple[int, str]] = [
+    (100, "Let me"),
+    (101, " think"),
+    (102, " about"),
+    (103, " Samsung."),
+    (51, "</think>"),
+    (200, "삼성"),
+    (201, "전자의"),
+    (202, " 주가를"),
+    (203, " 분석합니다."),
+]
+
+_SHARED_UFFFD_IDS: set[int] = {200}
+
+EXPECTED_REASONING = "Let me think about Samsung."
+EXPECTED_CONTENT = "삼성전자의 주가를 분석합니다."
+
+_MODEL_CONFIGS = [
+    pytest.param(
+        {
+            "<think>": 50,
+            "</think>": 51,
+            "<tool_call>": 60,
+            "</tool_call>": 61,
+            "<arg_key>": 62,
+            "</arg_key>": 63,
+            "<arg_value>": 64,
+            "</arg_value>": 65,
+        },
+        _Glm47Delegating,
+        id="glm47",
+    ),
+    pytest.param(
+        {
+            "<think>": 50,
+            "</think>": 51,
+            "<tool_call>": 60,
+            "</tool_call>": 61,
+        },
+        _Qwen3Delegating,
+        id="qwen3",
+    ),
+]
+
+
+# ── Tests ────────────────────────────────────────────────────────────
+
+
+class TestUfffdReasoningTransition:
+    """U+FFFD must not appear at the reasoning→content transition."""
+
+    @pytest.mark.parametrize("vocab,delegating_cls", _MODEL_CONFIGS)
+    @pytest.mark.parametrize("chunk_size", CHUNK_SIZES, ids=lambda c: f"chunk={c}")
+    def test_no_ufffd(self, chunk_size, vocab, delegating_cls):
+        tokenizer = ByteFallbackMockTokenizer(vocab, _SHARED_TOKENS, _SHARED_UFFFD_IDS)
+        parser = delegating_cls(tokenizer)
+        deltas = replay_streaming(
+            parser,
+            _SHARED_TOKENS,
+            chunk_size=chunk_size,
+            finished_on_last=True,
+        )
+        output = collect_output(deltas)
+
+        assert "�" not in output.content, (
+            f"U+FFFD leaked into content: {output.content!r}"
+        )
+        assert output.content == EXPECTED_CONTENT
+        assert output.reasoning == EXPECTED_REASONING
+
+    def test_byte_fallback_tokenizer_produces_ufffd(self):
+        """Validate the fixture: decode() returns U+FFFD for isolated
+        byte-fallback token IDs, proving the old code path would leak."""
+        vocab = dict(_MODEL_CONFIGS[0].values[0])
+        tokenizer = ByteFallbackMockTokenizer(vocab, _SHARED_TOKENS, _SHARED_UFFFD_IDS)
+        assert tokenizer.decode([200]) == "�"
+
+    @pytest.mark.parametrize("chunk_size", CHUNK_SIZES, ids=lambda c: f"chunk={c}")
+    def test_multiple_ufffd_tokens_at_boundary(self, chunk_size):
+        """Multiple consecutive byte-fallback tokens at the boundary."""
+        tokens: list[tuple[int, str]] = [
+            (100, "Reasoning."),
+            (51, "</think>"),
+            (200, "삼"),
+            (201, "성"),
+            (202, "전자"),
+        ]
+        ufffd_ids: set[int] = {200, 201}
+        vocab = dict(_MODEL_CONFIGS[0].values[0])
+
+        tokenizer = ByteFallbackMockTokenizer(vocab, tokens, ufffd_ids)
+        parser = _Glm47Delegating(tokenizer)
+        deltas = replay_streaming(
+            parser,
+            tokens,
+            chunk_size=chunk_size,
+            finished_on_last=True,
+        )
+        output = collect_output(deltas)
+
+        assert "�" not in output.content, (
+            f"U+FFFD leaked into content: {output.content!r}"
+        )
+        assert output.content == "삼성전자"
+        assert output.reasoning == "Reasoning."
@@ -143,6 +143,12 @@ SCENARIOS: list[Scenario] = [
        tool_calls=[_READ_TOOL],
        after_tool_response=True,
    ),
+    Scenario(
+        id="empty-tool-block",
+        description="Empty tool block followed by content (edge case recovery)",
+        content="Content after empty tools.",
+        tool_calls=[],
+    ),
 ]


@@ -344,8 +350,11 @@ def _qwen3_segments(scenario: Scenario) -> list[tuple[str, bool]]:
    segs: list[tuple[str, bool]] = []
    if scenario.reasoning is not None:
        segs.append((scenario.reasoning, False))
-    if scenario.content is not None or scenario.tool_calls:
+    if scenario.content is not None or scenario.tool_calls is not None:
        segs.append(("</think>", True))
+    if scenario.tool_calls is not None and not scenario.tool_calls:
+        segs.append(("<tool_call>", True))
+        segs.append(("</tool_call>", True))
    if scenario.content is not None:
        segs.append((scenario.content, False))
    if scenario.tool_calls:
@@ -437,8 +446,11 @@ def _minimax_m2_segments(scenario: Scenario) -> list[tuple[str, bool]]:
    segs: list[tuple[str, bool]] = []
    if scenario.reasoning is not None:
        segs.append((scenario.reasoning, False))
-    if scenario.content is not None or scenario.tool_calls:
+    if scenario.content is not None or scenario.tool_calls is not None:
        segs.append(("</think>", True))
+    if scenario.tool_calls is not None and not scenario.tool_calls:
+        segs.append(("<minimax:tool_call>", True))
+        segs.append(("</minimax:tool_call>", True))
    if scenario.content is not None:
        segs.append((scenario.content, False))
    if scenario.tool_calls:
@@ -534,6 +546,9 @@ def _gemma4_segments(scenario: Scenario) -> list[tuple[str, bool]]:
        segs.append((_GEMMA4_THOUGHT_PREFIX, False))
        segs.append((scenario.reasoning, False))
        segs.append(("<channel|>", True))
+    if scenario.tool_calls is not None and not scenario.tool_calls:
+        segs.append(("<|tool_call>", True))
+        segs.append(("<tool_call|>", True))
    if scenario.content is not None:
        segs.append((scenario.content, False))
    if scenario.tool_calls:
@@ -18,6 +18,7 @@ from vllm.model_executor.kernels.linear.scaled_mm.ScaledMMLinearKernel import (
    FP8ScaledMMLinearLayerConfig,
 )
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    get_fp8_min_max,
    kFp8DynamicTokenSym,
    kFp8StaticChannelSym,
    kFp8StaticTensorSym,
@@ -309,7 +310,7 @@ def test_hipb_mm_kernel_forward_accuracy(enable_hipb_mm_kernel):
    _check_bpreshuffle_runtime_support(weight_shape, num_tokens=num_tokens)

    fp8_dtype = current_platform.fp8_dtype()
-    fp8_max = torch.finfo(fp8_dtype).max
+    fp8_max = get_fp8_min_max()[1]
    device = torch.device("cuda")

    # Build a bf16 weight and quantize per output channel (one scale per row).
@@ -0,0 +1,135 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Tests for contiguous KV cache packing in _get_kv_cache_config_deepseek_v4."""
+
+from unittest.mock import MagicMock
+
+import pytest
+import torch
+
+from vllm.v1.core.kv_cache_utils import _get_kv_cache_config_deepseek_v4
+from vllm.v1.kv_cache_interface import (
+    KVCacheGroupSpec,
+    MLAAttentionSpec,
+    UniformTypeKVCacheSpecs,
+)
+
+
+def _make_mla_spec(page_size: int, block_size: int = 256) -> MLAAttentionSpec:
+    return MLAAttentionSpec(
+        block_size=block_size,
+        num_kv_heads=1,
+        head_size=512,
+        dtype=torch.uint8,
+        page_size_padded=page_size,
+        cache_dtype_str="fp8_ds_mla",
+        model_version="deepseek_v4",
+        alignment=576,
+    )
+
+
+def _make_groups(n_c4, n_c128, n_swa):
+    PS_C4_MLA = 37440
+    PS_C4_IDX = 8640
+    PS_C128 = 1728
+    PS_SWA = 37440
+
+    mla_specs = {}
+    for i in range(n_c4):
+        mla_specs[f"c4_mla.{i}"] = _make_mla_spec(PS_C4_MLA)
+        mla_specs[f"c4_idx.{i}"] = _make_mla_spec(PS_C4_IDX)
+    for i in range(n_c128):
+        mla_specs[f"c128_mla.{i}"] = _make_mla_spec(PS_C128)
+
+    mla_group = KVCacheGroupSpec(
+        layer_names=list(mla_specs.keys()),
+        kv_cache_spec=UniformTypeKVCacheSpecs(block_size=256, kv_cache_specs=mla_specs),
+    )
+
+    swa_specs = {}
+    for i in range(n_swa):
+        swa_specs[f"swa.{i}"] = _make_mla_spec(PS_SWA)
+
+    swa_group = KVCacheGroupSpec(
+        layer_names=list(swa_specs.keys()),
+        kv_cache_spec=UniformTypeKVCacheSpecs(block_size=256, kv_cache_specs=swa_specs),
+    )
+
+    return [mla_group, swa_group]
+
+
+def _mock_vllm_config():
+    config = MagicMock()
+    config.cache_config.num_gpu_blocks_override = None
+    return config
+
+
+def _run(n_c4=3, n_c128=2, n_swa=5, mem=100 * 1024 * 1024):
+    groups = _make_groups(n_c4, n_c128, n_swa)
+    return _get_kv_cache_config_deepseek_v4(_mock_vllm_config(), groups, mem)
+
+
+def _page_sizes_by_layer(
+    groups: list[KVCacheGroupSpec],
+) -> dict[str, int]:
+    page_sizes = {}
+    for group in groups:
+        specs = group.kv_cache_spec.kv_cache_specs
+        for layer_name in group.layer_names:
+            page_sizes[layer_name] = specs[layer_name].page_size_bytes
+    return page_sizes
+
+
+class TestInterleavedPacking:
+    def test_all_tensors_have_block_stride(self):
+        _, tensors = _run()
+        for t in tensors:
+            assert t.block_stride > 0
+
+    def test_all_tensors_share_same_size(self):
+        _, tensors = _run()
+        sizes = set(t.size for t in tensors)
+        assert len(sizes) == 1
+        assert sizes.pop() > 0
+
+    def test_offsets_within_one_block(self):
+        _, tensors = _run()
+        for t in tensors:
+            assert t.offset < t.block_stride
+
+    def test_all_layers_accounted_for(self):
+        n_c4, n_c128, n_swa = 5, 4, 7
+        _, tensors = _run(n_c4=n_c4, n_c128=n_c128, n_swa=n_swa)
+        all_names = set()
+        for t in tensors:
+            all_names.update(t.shared_by)
+        expected = n_c4 * 2 + n_c128 + n_swa
+        assert len(all_names) == expected
+
+    def test_strided_views_are_independent(self):
+        groups = _make_groups(n_c4=3, n_c128=2, n_swa=5)
+        page_sizes = _page_sizes_by_layer(groups)
+        num_blocks, tensors = _get_kv_cache_config_deepseek_v4(
+            _mock_vllm_config(), groups, 100 * 1024 * 1024
+        )
+        backing = torch.zeros(tensors[0].size, dtype=torch.uint8)
+        views = []
+        for t in tensors:
+            page_size = page_sizes[t.shared_by[0]]
+            v = torch.as_strided(
+                backing,
+                size=(num_blocks, page_size),
+                stride=(t.block_stride, 1),
+                storage_offset=t.offset,
+            )
+            views.append(v)
+
+        for i, v in enumerate(views):
+            v.fill_(i + 1)
+
+        for i, v in enumerate(views):
+            assert (v == i + 1).all(), f"View {i} was corrupted"
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])
@@ -4,9 +4,16 @@
 import pytest

 from vllm import LLM, SamplingParams
+from vllm.platforms import current_platform

 from ....utils import create_new_process_for_each_test

+if current_platform.is_rocm():
+    pytest.skip(
+        "Cascade attention backends FLASH_ATTN and FLASHINFER are notsupported on ROCm",
+        allow_module_level=True,
+    )
+

@create_new_process_for_each_test()
@pytest.mark.parametrize("attn_backend", ["FLASH_ATTN", "FLASHINFER"])
@@ -1,6 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

+import threading
+import time
 from unittest.mock import MagicMock, patch

 from vllm.config import set_current_vllm_config
@@ -406,7 +408,9 @@ def test_lookup_key_client_lookup_prepends_typed_tag():
    fake_socket = mock_make_socket.return_value
    fake_socket.recv.return_value = (5).to_bytes(4, "big")

-    assert client.lookup(token_len=128, block_hashes=[]) == 5
+    # Blocking lookup (non_block defaults to False) runs on the executor and
+    # returns the resolved hit length.
+    assert client.lookup("req0", token_len=128, block_hashes=[]) == 5

    sent_frames = fake_socket.send_multipart.call_args[0][0]
    assert sent_frames[0] == protocol.LOOKUP_MSG
@@ -435,6 +439,127 @@ def test_lookup_key_client_reset_uses_typed_protocol():
    assert client.reset() is False


+def _poll_lookup(client, req_id, token_len=128, block_hashes=(), timeout=5.0):
+    """Drive non-blocking lookup until the executor completes it."""
+    deadline = time.monotonic() + timeout
+    while time.monotonic() < deadline:
+        result = client.lookup(req_id, token_len, list(block_hashes), non_block=True)
+        if result is not None:
+            return result
+        time.sleep(0.005)
+    return None
+
+
+def _gated_recv(gate: threading.Event, value: int):
+    """Mock recv side-effect that blocks until ``gate`` is set, so the
+    executor's lookup can be held pending deterministically."""
+
+    def recv():
+        gate.wait()
+        return value.to_bytes(4, "big")
+
+    return recv
+
+
+def test_lookup_key_client_non_block_lookup_async():
+    """Non-blocking lookup defers to the executor: None first, hit once the
+    Future resolves."""
+    vllm_config = _make_vllm_config()
+
+    with patch(
+        "vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store."
+        "worker.make_zmq_socket"
+    ) as mock_make_socket:
+        client = worker.LookupKeyClient(vllm_config)
+
+    fake_socket = mock_make_socket.return_value
+    # Hold the executor's lookup pending until we release the gate.
+    gate = threading.Event()
+    fake_socket.recv.side_effect = _gated_recv(gate, 7)
+
+    # First query submits the lookup and returns None while it is in flight.
+    assert client.lookup("req1", 128, [], non_block=True) is None
+    # Release the executor; a later poll returns the hit length.
+    gate.set()
+    assert _poll_lookup(client, "req1") == 7
+    # Future is consumed (popped) on read.
+    assert "req1" not in client.futures
+
+
+def test_lookup_key_client_discard_clears_state():
+    """discard() drops a completed lookup Future so it is not served stale."""
+    vllm_config = _make_vllm_config()
+
+    with patch(
+        "vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store."
+        "worker.make_zmq_socket"
+    ) as mock_make_socket:
+        client = worker.LookupKeyClient(vllm_config)
+
+    fake_socket = mock_make_socket.return_value
+    gate = threading.Event()
+    fake_socket.recv.side_effect = _gated_recv(gate, 9)
+
+    # Submit while gated so the call returns None and the Future stays in
+    # `futures` (unconsumed) once it resolves.
+    assert client.lookup("req2", 128, [], non_block=True) is None
+    gate.set()
+    deadline = time.monotonic() + 5.0
+    while time.monotonic() < deadline:
+        if client.futures["req2"].done():
+            break
+        time.sleep(0.005)
+    # discard() drops the completed result before any lookup consumes it.
+    client.discard("req2")
+    assert "req2" not in client.futures
+    # A fresh query re-submits rather than returning a stale value: hold the
+    # gate so the resubmitted lookup stays in flight.
+    gate.clear()
+    assert client.lookup("req2", 128, [], non_block=True) is None
+    gate.set()  # release the executor so the worker thread can drain
+
+
+def test_get_num_new_matched_tokens_async_defers_then_reports():
+    """Async lookup returns (None, False) until ready, then the hit count."""
+    vllm_config = create_vllm_config(
+        kv_connector="MooncakeStoreConnector",
+        kv_role="kv_both",
+        kv_connector_extra_config={"lookup_async": True},
+    )
+    kv_cache_config = _make_kv_cache_config()
+
+    with (
+        set_current_vllm_config(vllm_config),
+        patch(
+            "vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store."
+            "scheduler.LookupKeyClient"
+        ) as mock_client_cls,
+    ):
+        sched = scheduler.MooncakeStoreScheduler(vllm_config, kv_cache_config)
+
+    assert sched.lookup_async is True
+    mock_client = mock_client_cls.return_value
+
+    block_size = sched._block_size
+    request = MagicMock()
+    request.request_id = "r1"
+    request.num_tokens = 4 * block_size
+    request.block_hashes = []
+
+    # Lookup not ready -> defer.
+    mock_client.lookup.return_value = None
+    assert sched.get_num_new_matched_tokens(request, 0) == (None, False)
+    assert "r1" not in sched.load_specs
+
+    # Lookup ready with a hit -> report need_to_allocate + async-load flag.
+    hit = 3 * block_size
+    mock_client.lookup.return_value = hit
+    need, load_async = sched.get_num_new_matched_tokens(request, 0)
+    assert need == hit
+    assert load_async == sched.load_async
+    assert sched.load_specs["r1"].kvpool_cached_tokens == hit
+
+
 def test_protocol_tags_are_distinct_and_non_empty():
    """Protocol tags must be unique and non-empty to avoid collision."""
    tags = {protocol.LOOKUP_MSG, protocol.RESET_MSG}
@@ -16,6 +16,7 @@ from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store.scheduler impor
 def _make_bare_scheduler() -> MooncakeStoreScheduler:
    scheduler = object.__new__(MooncakeStoreScheduler)
    scheduler.kv_role = "kv_both"
+    scheduler.lookup_async = False
    scheduler._block_size = 16
    scheduler.load_specs = {}
    scheduler._preempted_req_ids = set()
@@ -405,7 +406,13 @@ class _StubLookupClient:
    def __init__(self, hit_tokens: int) -> None:
        self._hit_tokens = hit_tokens

-    def lookup(self, token_len: int, block_hashes: list[bytes]) -> int:
+    def lookup(
+        self,
+        req_id: str,
+        token_len: int,
+        block_hashes: list[bytes],
+        non_block: bool = False,
+    ) -> int:
        return self._hit_tokens


@@ -175,14 +175,17 @@ class _FakeModelConfig:


 def _make_vllm_config(
-    *, extra_config: dict[str, object] | None = None
+    *,
+    extra_config: dict[str, object] | None = None,
+    rank: int = 0,
+    decode_context_parallel_size: int = 1,
 ) -> SimpleNamespace:
    return SimpleNamespace(
        model_config=_FakeModelConfig(),
        parallel_config=SimpleNamespace(
            pipeline_parallel_size=1,
-            rank=0,
-            decode_context_parallel_size=1,
+            rank=rank,
+            decode_context_parallel_size=decode_context_parallel_size,
            prefill_context_parallel_size=1,
        ),
        kv_transfer_config=_FakeKVTransferConfig(extra_config=extra_config),
@@ -231,13 +234,23 @@ def _install_fake_mooncake(monkeypatch, store_instance: MagicMock):
    return FakeReplicateConfig


-def _patch_worker_runtime(monkeypatch, *, local_ip: str = "10.0.0.7") -> None:
+def _patch_worker_runtime(
+    monkeypatch,
+    *,
+    local_ip: str = "10.0.0.7",
+    tp_rank: int = 0,
+    tp_size: int = 1,
+    dcp_size: int = 1,
+) -> None:
    single_rank_group = SimpleNamespace(world_size=1, rank_in_group=0)
+    # DCP groups are contiguous splits of the TP group (see
+    # parallel_state.py), so dcp_rank == tp_rank % dcp_size.
+    dcp_group = SimpleNamespace(world_size=dcp_size, rank_in_group=tp_rank % dcp_size)
    monkeypatch.setattr(worker, "get_mooncake_dp_engine_index", lambda _: 0)
-    monkeypatch.setattr(worker, "get_tensor_model_parallel_rank", lambda: 0)
-    monkeypatch.setattr(worker, "get_tensor_model_parallel_world_size", lambda: 1)
+    monkeypatch.setattr(worker, "get_tensor_model_parallel_rank", lambda: tp_rank)
+    monkeypatch.setattr(worker, "get_tensor_model_parallel_world_size", lambda: tp_size)
    monkeypatch.setattr(worker, "get_pcp_group", lambda: single_rank_group)
-    monkeypatch.setattr(worker, "get_dcp_group", lambda: single_rank_group)
+    monkeypatch.setattr(worker, "get_dcp_group", lambda: dcp_group)
    monkeypatch.setattr(worker, "get_ip", lambda: local_ip)


@@ -884,6 +897,66 @@ def test_requester_worker_init_builds_replicate_config_for_preferred_segment(
    assert w.store_replicate_config.preferred_segment == "10.0.0.7:50053"


+@pytest.mark.parametrize("dcp_size", [1, 4])
+def test_worker_put_striding_covers_every_rank_get_namespace(
+    tmp_path, monkeypatch, dcp_size
+):
+    """Every key a rank GETs must have been PUT by some rank.
+
+    When num_kv_head < tp_size, ranks holding the same KV heads stripe
+    their PUTs across one shared key namespace. That dedup is only valid
+    when those ranks really share a namespace: with DCP > 1 each rank GETs
+    every key from its own ``@dcpN`` namespace, so striding must be
+    disabled.
+    """
+    tp_size = 4
+    store = MagicMock()
+    store.setup.return_value = 0
+    _install_fake_mooncake(monkeypatch, store)
+    monkeypatch.setenv(
+        "MOONCAKE_CONFIG_PATH",
+        _write_mooncake_config(
+            tmp_path,
+            {
+                "metadata_server": "http://metadata/endpoint",
+                "protocol": "tcp",
+                "device_name": "",
+                "master_server_address": "10.0.0.7:50051",
+            },
+        ),
+    )
+
+    # _FakeModelConfig has num_kv_head=1 < tp_size, which enables striding.
+    block_hashes = [f"hash-{i}".encode() for i in range(4)]
+    put_keys: set[str] = set()
+    get_keys_per_rank: dict[int, set[str]] = {}
+    for tp_rank in range(tp_size):
+        _patch_worker_runtime(
+            monkeypatch, tp_rank=tp_rank, tp_size=tp_size, dcp_size=dcp_size
+        )
+        w = worker.MooncakeStoreWorker(
+            _make_vllm_config(rank=tp_rank, decode_context_parallel_size=dcp_size),
+            _make_kv_cache_config(),
+        )
+        db = w.token_dbs[0]
+        token_len = len(block_hashes) * db.block_size
+        keys = [
+            key.to_string() for _, _, key in db.process_tokens(token_len, block_hashes)
+        ]
+        assert len(keys) == len(block_hashes)
+        # PUT side: mirrors KVCacheStoreSendingThread's striding slice.
+        put_keys.update(keys[w.tp_rank % w.put_step :: w.put_step])
+        # GET side: KVCacheStoreRecvingThread fetches every key.
+        get_keys_per_rank[tp_rank] = set(keys)
+
+    for tp_rank, rank_keys in get_keys_per_rank.items():
+        missing = rank_keys - put_keys
+        assert not missing, (
+            f"tp_rank={tp_rank} would GET {len(missing)}/{len(rank_keys)} keys "
+            f"that no rank PUT (Mooncake OBJECT_NOT_FOUND): {sorted(missing)}"
+        )
+
+
 # ---------------------------------------------------------------------------
 # Helpers for register_kv_caches tests
 # ---------------------------------------------------------------------------
@@ -0,0 +1,227 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Tests for the Model Runner V2 Gumbel-max sampling kernel.
+
+Accuracy: define a target categorical distribution as a non-negative int64
+count tensor summing to N, turn it into logits (= log(count)), sample many
+times with `gumbel_sample`, and check the empirical distribution matches.
+
+The count tensor is deliberately heavy-tailed (one dominant token, the rest
+~18 logits below). That tail is the sensitive part: the fp32 Gumbel noise must
+reach ~18 to ever sample it. A flat distribution would keep every token within
+a few logits of the top and would not exercise the noise tail at all.
+"""
+
+import math
+
+import pytest
+import torch
+
+pytest.importorskip("triton")
+if not torch.cuda.is_available():
+    pytest.skip("CUDA required for Gumbel sampler tests", allow_module_level=True)
+
+from vllm.v1.worker.gpu.sample.gumbel import gumbel_sample
+
+DEVICE = "cuda"
+VOCAB_SIZE = 200_000
+NUM_SAMPLES = 500_000
+# Dominant token is exp(HEAD_LOG_GAP)x larger than the unit-count tail, so the
+# tail sits ~HEAD_LOG_GAP logits below the top.
+HEAD_LOG_GAP = 18.0
+# 10-sigma band: a correct sampler effectively never trips it.
+Z_TOLERANCE = 10.0
+
+
+def _make_heavy_tailed_counts(seed: int = 1234) -> torch.Tensor:
+    """Non-negative int64 counts of shape [VOCAB_SIZE]; target prob = counts/N."""
+    gen = torch.Generator(device=DEVICE).manual_seed(seed)
+    counts = torch.randint(
+        1, 4, (VOCAB_SIZE,), generator=gen, dtype=torch.int64, device=DEVICE
+    )
+    counts[0] = round(math.exp(HEAD_LOG_GAP))  # dominant token
+    return counts
+
+
+def _counts_to_logits(counts: torch.Tensor) -> torch.Tensor:
+    # softmax(log(count)) == count / sum(count); count 0 -> logit -inf -> prob 0.
+    return counts.double().log().to(torch.float32)
+
+
+def _sample(
+    logits_1d: torch.Tensor,
+    num_samples: int,
+    *,
+    use_fp64: bool = False,
+    temperature: float = 1.0,
+) -> torch.Tensor:
+    """Sample `num_samples` tokens from one logit vector.
+
+    Fixed seed with a distinct `pos` per sample gives independent draws; the
+    logits are broadcast with a 0-stride view to avoid materializing
+    [num_samples, vocab_size].
+    """
+    vocab_size = logits_1d.shape[0]
+    logits = logits_1d.unsqueeze(0).expand(num_samples, vocab_size)
+    idx_mapping = torch.zeros(num_samples, dtype=torch.int32, device=DEVICE)
+    temp = torch.tensor([temperature], dtype=torch.float32, device=DEVICE)
+    seed = torch.tensor([0xABCD], dtype=torch.int64, device=DEVICE)
+    pos = torch.arange(num_samples, dtype=torch.int64, device=DEVICE)
+    return gumbel_sample(
+        logits,
+        idx_mapping,
+        temp,
+        seed,
+        pos,
+        apply_temperature=True,
+        use_fp64=use_fp64,
+    )
+
+
+def _z_score(observed: int, expected: float, num_trials: int) -> float:
+    p = expected / num_trials
+    return (observed - expected) / math.sqrt(num_trials * p * (1 - p))
+
+
+def _sample_histogram(
+    logits_1d: torch.Tensor, num_samples: int, *, chunk: int = 1_000_000
+) -> torch.Tensor:
+    """Histogram of `num_samples` draws, accumulated in chunks.
+
+    Chunking keeps the kernel's per-sample scratch ([chunk, num_blocks]) bounded
+    so a large sample count does not blow up memory.
+    """
+    vocab_size = logits_1d.shape[0]
+    hist = torch.zeros(vocab_size, dtype=torch.float64, device=DEVICE)
+    for start in range(0, num_samples, chunk):
+        size = min(chunk, num_samples - start)
+        logits = logits_1d.unsqueeze(0).expand(size, vocab_size)
+        idx_mapping = torch.zeros(size, dtype=torch.int32, device=DEVICE)
+        temp = torch.tensor([1.0], dtype=torch.float32, device=DEVICE)
+        seed = torch.tensor([0xABCD], dtype=torch.int64, device=DEVICE)
+        pos = torch.arange(start, start + size, dtype=torch.int64, device=DEVICE)
+        out = gumbel_sample(
+            logits, idx_mapping, temp, seed, pos, apply_temperature=True
+        )
+        hist += torch.bincount(out, minlength=vocab_size).double()
+    return hist
+
+
+# ----------------------------- Accuracy ------------------------------------
+
+
+@pytest.mark.parametrize("use_fp64", [False, True])
+def test_sampling_matches_target_distribution(use_fp64: bool):
+    counts = _make_heavy_tailed_counts()
+    total = counts.sum().item()
+    logits = _counts_to_logits(counts)
+
+    sampled = _sample(logits, NUM_SAMPLES, use_fp64=use_fp64)
+    assert sampled.min() >= 0 and sampled.max() < VOCAB_SIZE
+
+    # The dominant token (index 0) and the aggregate tail are the two
+    # statistically resolvable bins (individual tail tokens are far below the
+    # ~5/N detectability floor). The tail mass is small but well above noise,
+    # and it lives beyond the fp32 Gumbel cap -- the regime sensitive to noise
+    # precision -- so matching it is the meaningful check.
+    tail_prob = (total - counts[0].item()) / total
+    tail_count = (sampled != 0).sum().item()
+    z = _z_score(tail_count, NUM_SAMPLES * tail_prob, NUM_SAMPLES)
+    assert abs(z) < Z_TOLERANCE, (
+        f"sampled tail mass {tail_count / NUM_SAMPLES:.3e} != target "
+        f"{tail_prob:.3e} (z={z:.2f})"
+    )
+
+
+def test_full_vocab_distribution_fidelity():
+    """The sampled distribution matches the target across the WHOLE vocab.
+
+    A near-flat count tensor makes every one of the 200K bins individually
+    measurable. With ~20 samples/bin, a goodness-of-fit over all bins checks
+    that no part of the vocab is over- or under-represented (the heavy-tailed
+    test above only resolves head vs aggregate tail). Empirically the fp32
+    sampler is as faithful here as torch.multinomial; the residual error is the
+    multinomial sampling-noise floor, not the kernel.
+    """
+    gen = torch.Generator(device=DEVICE).manual_seed(2024)
+    counts = torch.randint(
+        500, 1500, (VOCAB_SIZE,), generator=gen, dtype=torch.int64, device=DEVICE
+    )
+    total = counts.sum().item()
+    logits = _counts_to_logits(counts)
+
+    num_samples = 4_000_000
+    hist = _sample_histogram(logits, num_samples)
+
+    # Diversity: essentially every token must be reachable (no starved region).
+    coverage = (hist > 0).sum().item() / VOCAB_SIZE
+    assert coverage > 0.99, f"only {coverage:.4f} of the vocab was ever sampled"
+
+    # Goodness-of-fit across all bins (each has expected count >= ~10).
+    expected = (counts.double() / total) * num_samples
+    chi2 = (((hist - expected) ** 2) / expected).sum().item()
+    df = VOCAB_SIZE - 1
+    assert chi2 < df + 10 * math.sqrt(2 * df), f"chi2={chi2:.0f}, df={df}"
+
+
+# ----------------------------- Edge cases ----------------------------------
+
+
+def test_greedy_temperature_zero_returns_argmax():
+    """temperature == 0 skips Gumbel noise and returns the exact argmax."""
+    torch.manual_seed(0)
+    num_reqs = 128
+    logits = torch.randn(num_reqs, VOCAB_SIZE, device=DEVICE, dtype=torch.float32)
+    idx_mapping = torch.arange(num_reqs, dtype=torch.int32, device=DEVICE)
+    temp = torch.zeros(num_reqs, dtype=torch.float32, device=DEVICE)
+    seed = torch.arange(num_reqs, dtype=torch.int64, device=DEVICE)
+    pos = torch.arange(num_reqs, dtype=torch.int64, device=DEVICE)
+
+    sampled = gumbel_sample(
+        logits, idx_mapping, temp, seed, pos, apply_temperature=True
+    )
+    assert torch.equal(sampled, logits.argmax(dim=-1))
+
+
+def test_zero_count_tokens_are_never_sampled():
+    """Count 0 -> -inf logit -> probability 0; must never be selected."""
+    counts = _make_heavy_tailed_counts(seed=7)
+    zeroed = torch.arange(1, VOCAB_SIZE, 2, device=DEVICE)  # odd indices (not head)
+    counts[zeroed] = 0
+    logits = _counts_to_logits(counts)
+
+    sampled = _sample(logits, NUM_SAMPLES)
+    assert sampled.min() >= 0 and sampled.max() < VOCAB_SIZE
+    assert not torch.isin(sampled, zeroed).any(), "sampled a zero-probability token"
+
+
+def test_single_nonzero_token_is_always_sampled():
+    """A lone finite logit must win every draw, regardless of its index."""
+    counts = torch.zeros(VOCAB_SIZE, dtype=torch.int64, device=DEVICE)
+    counts[123_456] = 1000
+    logits = _counts_to_logits(counts)
+
+    sampled = _sample(logits, 10_000)
+    assert (sampled == 123_456).all()
+
+
+@pytest.mark.parametrize("vocab_size", [1, 999, 1024, 4097])
+def test_vocab_size_not_multiple_of_block(vocab_size: int):
+    """Per-block tail masking for non-block-aligned vocab; all bins measurable."""
+    gen = torch.Generator(device=DEVICE).manual_seed(vocab_size)
+    counts = torch.randint(
+        20, 200, (vocab_size,), generator=gen, dtype=torch.int64, device=DEVICE
+    )
+    total = counts.sum().item()
+    logits = _counts_to_logits(counts)
+    num_samples = max(40 * vocab_size, 50_000)
+
+    sampled = _sample(logits, num_samples)
+    assert sampled.min() >= 0 and sampled.max() < vocab_size
+
+    observed = torch.bincount(sampled, minlength=vocab_size).double()
+    expected = (counts.double() / total) * num_samples
+    chi2 = (((observed - expected) ** 2) / expected).sum().item()
+    df = vocab_size - 1
+    if df >= 1:
+        assert chi2 < df + 10 * math.sqrt(2 * df), f"chi2={chi2:.1f}, df={df}"
@@ -991,7 +991,7 @@ class VllmBackend:
            },
            payload_fn=lambda: json.dumps(
                {
-                    "model": self.vllm_config.model_config.model,
+                    "model": getattr(self.vllm_config.model_config, "model", "unknown"),
                    "prefix": self.prefix,
                    "mode": str(cc.mode),
                    "backend": cc.backend,
@@ -302,6 +302,14 @@ class ParallelConfig:
    Each entry must use `numactl --physcpubind` CPU-list syntax, for example
    `"0-3"` or `"0,2,4-7"`.
    """
+    assigned_physical_gpu_ids: list[int] | None = None
+    """Mapping from vLLM-local logical GPU IDs to physical GPU IDs.
+
+    For example, ``[2, 3]`` means logical GPU 0 maps to physical GPU 2,
+    and logical GPU 1 maps to physical GPU 3. Physical IDs are used only
+    at platform/topology boundaries such as NVML, NIC affinity, P2P
+    checks, and final CUDA device selection when needed. When None,
+    logical IDs map to visible device IDs in order."""

    distributed_timeout_seconds: int | None = None
    """Timeout in seconds for distributed operations (e.g., init_process_group).
@@ -772,6 +780,7 @@ class ParallelConfig:
            "numa_bind",
            "numa_bind_nodes",
            "numa_bind_cpus",
+            "assigned_physical_gpu_ids",
        }

        from vllm.config.utils import get_hash_factors, hash_factors
@@ -704,7 +704,14 @@ class FlashInferNVLinkOneSidedManager(All2AllManagerBase):
        self.num_experts = num_experts

        self.cleanup()
-        gpus_per_node = torch.accelerator.device_count()
+        from vllm.platforms.interface import get_assigned_physical_gpu_ids
+
+        assigned_physical_gpu_ids = get_assigned_physical_gpu_ids()
+        gpus_per_node = (
+            len(assigned_physical_gpu_ids)
+            if assigned_physical_gpu_ids is not None
+            else torch.accelerator.device_count()
+        )
        logger.debug(
            "Making One-sided NVLink mapping: rank=%d, world size=%d",
            self.rank,
@@ -320,13 +320,21 @@ def gpu_p2p_access_check(src: int, tgt: int) -> bool:

    is_distributed = dist.is_initialized()

-    num_dev = current_platform.device_count()
-    cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES
-    if cuda_visible_devices is None:
-        cuda_visible_devices = ",".join(str(i) for i in range(num_dev))
+    from vllm.platforms.interface import get_assigned_physical_gpu_ids
+
+    assigned_physical_gpu_ids = get_assigned_physical_gpu_ids()
+    if assigned_physical_gpu_ids is not None:
+        # Key by the ordered list: the cache stores directed local-index
+        # pairs, so permutations of the same set are distinct mappings.
+        cache_key = ",".join(str(i) for i in assigned_physical_gpu_ids)
+        num_dev = len(assigned_physical_gpu_ids)
+    else:
+        num_dev = current_platform.device_count()
+        cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES
+        cache_key = cuda_visible_devices or ",".join(str(i) for i in range(num_dev))

    path = os.path.join(
-        envs.VLLM_CACHE_ROOT, f"gpu_p2p_access_cache_for_{cuda_visible_devices}.json"
+        envs.VLLM_CACHE_ROOT, f"gpu_p2p_access_cache_for_{cache_key}.json"
    )
    os.makedirs(os.path.dirname(path), exist_ok=True)
    from vllm.distributed.parallel_state import get_world_group
@@ -338,7 +346,15 @@ def gpu_p2p_access_check(src: int, tgt: int) -> bool:
        #  enter this block to calculate the cache
        logger.info("generating GPU P2P access cache in %s", path)
        cache: dict[str, bool] = {}
-        ids = list(range(num_dev))
+        # The probe subprocesses inherit this process's device-control env
+        # var, so they must be given visible ordinals, not physical IDs.
+        if assigned_physical_gpu_ids is not None:
+            ids = [
+                current_platform.logical_device_id_to_visible_device_id(local)
+                for local in range(num_dev)
+            ]
+        else:
+            ids = list(range(num_dev))
        # batch of all pairs of GPUs
        batch_src, batch_tgt = zip(*list(product(ids, ids)))
        # NOTE: we use `subprocess` rather than `multiprocessing` here
@@ -368,8 +384,11 @@ def gpu_p2p_access_check(src: int, tgt: int) -> bool:
                ) from e
            with open(output_file.name, "rb") as f:
                result = pickle.load(f)
+        # Cache entries must be keyed by local indices (0..N-1) because
+        # gpu_p2p_access_check() is called with local ranks.
+        id_to_local = {device_id: local for local, device_id in enumerate(ids)}
        for _i, _j, r in zip(batch_src, batch_tgt, result):
-            cache[f"{_i}->{_j}"] = r
+            cache[f"{id_to_local[_i]}->{id_to_local[_j]}"] = r
        with open(path, "w") as f:
            json.dump(cache, f, indent=4)
    if is_distributed:
@@ -34,7 +34,12 @@ def _can_p2p(rank: int, world_size: int) -> bool:
            continue
        if envs.VLLM_SKIP_P2P_CHECK:
            logger.debug("Skipping P2P check and trusting the driver's P2P report.")
-            return torch.cuda.can_device_access_peer(rank, i)
+            # can_device_access_peer takes visible device ordinals, while
+            # rank and i are logical local IDs.
+            return torch.cuda.can_device_access_peer(
+                current_platform.logical_device_id_to_visible_device_id(rank),
+                current_platform.logical_device_id_to_visible_device_id(i),
+            )
        if not gpu_p2p_access_check(rank, i):
            return False
    return True
@@ -126,13 +131,10 @@ class CustomAllreduce:
                    CUSTOM_ALL_REDUCE_MAX_SIZES[device_capability_str][world_size],
                    max_size,
                )
-        cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES
-        if cuda_visible_devices:
-            device_ids = list(map(int, cuda_visible_devices.split(",")))
-        else:
-            device_ids = list(range(current_platform.device_count()))
-
-        physical_device_id = device_ids[device.index]
+        # device.index is a visible ordinal, not a logical local ID.
+        physical_device_id = current_platform.visible_device_id_to_physical_device_id(
+            device.index
+        )
        tensor = torch.tensor([physical_device_id], dtype=torch.int, device="cpu")
        gather_list = [
            torch.tensor([0], dtype=torch.int, device="cpu") for _ in range(world_size)
@@ -129,12 +129,10 @@ class QuickAllReduce:
        assert isinstance(device, torch.device)
        self.device = device

-        cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES
-        if cuda_visible_devices:
-            device_ids = list(map(int, cuda_visible_devices.split(",")))
-        else:
-            device_ids = list(range(current_platform.device_count()))
-        physical_device_id = device_ids[device.index]
+        # device.index is a visible ordinal, not a logical local ID.
+        physical_device_id = current_platform.visible_device_id_to_physical_device_id(
+            device.index
+        )
        tensor = torch.tensor([physical_device_id], dtype=torch.int, device="cpu")
        gather_list = [
            torch.tensor([0], dtype=torch.int, device="cpu")
@@ -840,7 +840,13 @@ class MessageQueue:
            The MessageQueue instance for the calling process,
            and a list of handles (only non-empty for the reader process).
        """
-        local_size = current_platform.device_count()
+        from vllm.platforms.interface import get_assigned_physical_gpu_ids
+
+        assigned_physical_gpu_ids = get_assigned_physical_gpu_ids()
+        if assigned_physical_gpu_ids is not None:
+            local_size = len(assigned_physical_gpu_ids)
+        else:
+            local_size = current_platform.device_count()
        rank = dist.get_rank()
        same_node = rank // local_size == reader_rank // local_size
        buffer_io = MessageQueue(
@@ -482,10 +482,11 @@ def _init_lmcache_engine(
    )

    # Change current device.
-    num_gpus = torch.accelerator.device_count()
-    local_rank = parallel_config.rank % num_gpus
-    torch.accelerator.set_device_index(local_rank)
-    device = torch.device(f"cuda:{local_rank}")
+    from vllm.distributed.parallel_state import get_world_group
+
+    device_index = get_world_group().device_index
+    torch.accelerator.set_device_index(device_index)
+    device = torch.device(f"cuda:{device_index}")
    metadata = LMCacheEngineMetadata(
        model_config.model,
        parallel_config.world_size,
@@ -176,7 +176,7 @@ class MooncakeStoreConnector(KVConnectorBase_V1, SupportsHMA):
        self,
        request: Request,
        num_computed_tokens: int,
-    ) -> tuple[int, bool]:
+    ) -> tuple[int | None, bool]:
        assert self.connector_scheduler is not None
        return self.connector_scheduler.get_num_new_matched_tokens(
            request, num_computed_tokens
@@ -54,9 +54,9 @@ class MooncakeStoreScheduler:
    ):
        assert vllm_config.kv_transfer_config is not None
        self.kv_role = vllm_config.kv_transfer_config.kv_role
-        self.load_async = vllm_config.kv_transfer_config.kv_connector_extra_config.get(
-            "load_async", True
-        )
+        kvc_extra_config = vllm_config.kv_transfer_config.kv_connector_extra_config
+        self.load_async = kvc_extra_config.get("load_async", True)
+        self.lookup_async = kvc_extra_config.get("lookup_async", False)
        self.client = LookupKeyClient(vllm_config)

        # Align with the engine's own scheduler_block_size and hash_block_size.
@@ -75,14 +75,26 @@ class MooncakeStoreScheduler:
        self,
        request: Request,
        num_computed_tokens: int,
-    ) -> tuple[int, bool]:
-        """Check for external KV cache hit."""
+    ) -> tuple[int | None, bool]:
+        """Check for external KV cache hit.
+
+        Returns ``(None, False)`` when an async lookup is still in flight,
+        signaling the scheduler to retry this request on a later step.
+        """
        # Look up against the full prefill range, not just the prompt.
        token_len = request.num_tokens // self._block_size * self._block_size
        if token_len < self._block_size:
            return 0, False

-        num_external_hit_tokens = self.client.lookup(token_len, request.block_hashes)
+        num_external_hit_tokens = self.client.lookup(
+            request.request_id,
+            token_len,
+            request.block_hashes,
+            non_block=self.lookup_async,
+        )
+        if num_external_hit_tokens is None:
+            # Lookup not ready yet; scheduler will retry on a later step.
+            return None, False

        if num_external_hit_tokens == request.num_tokens:
            # Leave a sub-block tail uncomputed for sampling, on a block
@@ -158,6 +170,7 @@ class MooncakeStoreScheduler:
        force_skip_save = self.kv_role == "kv_consumer"

        for finished_req_id in scheduler_output.finished_req_ids:
+            self.client.discard(finished_req_id)
            self.load_specs.pop(finished_req_id, None)
            self._request_trackers.pop(finished_req_id, None)
            self._unfinished_requests.pop(finished_req_id, None)
@@ -19,6 +19,7 @@ import threading
 import time
 from collections import defaultdict
 from collections.abc import Callable
+from concurrent.futures import Future, ThreadPoolExecutor
 from dataclasses import dataclass
 from typing import Any, Literal, TypeVar

@@ -971,7 +972,13 @@ class MooncakeStoreWorker:
        else:
            self.num_kv_head = model_config.get_total_num_kv_heads()

-        if self.num_kv_head < self.tp_size:
+        if self.num_kv_head < self.tp_size and self.dcp_size <= 1:
+            # Dedup: TP ranks holding the same KV heads stripe PUTs across
+            # one shared key namespace. DCP splits the TP group, so with
+            # DCP>1 those ranks have different `@dcpN` namespaces and
+            # striping would leave keys unwritten (OBJECT_NOT_FOUND on
+            # GET). PCP is outer to TP (pcp_rank is constant within a TP
+            # group), so it needs no guard.
            self.put_step = self.tp_size // self.num_kv_head
            self.head_or_tp_rank = self.tp_rank // self.put_step
        else:
@@ -1560,7 +1567,13 @@ class LookupKeyClient:
            bind=False,
        )

-    def lookup(self, token_len: int, block_hashes: list[BlockHash]) -> int:
+        # Async lookup support
+        self.executor = ThreadPoolExecutor(
+            max_workers=1, thread_name_prefix="MooncakeLookupClient"
+        )
+        self.futures: dict[str, Future[int]] = {}
+
+    def _lookup(self, token_len: int, block_hashes: list[BlockHash]) -> int:
        hash_strs = [h.hex() for h in block_hashes]
        hash_frames = self.encoder.encode(hash_strs)
        token_len_bytes = token_len.to_bytes(4, byteorder="big")
@@ -1570,7 +1583,36 @@ class LookupKeyClient:
        result = int.from_bytes(resp, "big")
        return result

-    def reset(self) -> bool:
+    def lookup(
+        self,
+        req_id: str,
+        token_len: int,
+        block_hashes: list[BlockHash],
+        non_block: bool = False,
+    ) -> int | None:
+        """If non_block is True, will return None until the result is ready,
+        so the caller retries on a later step."""
+        future = self.futures.get(req_id)
+        if future is None:
+            future = self.executor.submit(self._lookup, token_len, list(block_hashes))
+            self.futures[req_id] = future
+        if non_block and not future.done():
+            return None
+        try:
+            return future.result()
+        except Exception as e:
+            logger.error("Async Mooncake lookup failed for %s: %s", req_id, e)
+            return 0
+        finally:
+            del self.futures[req_id]
+
+    def discard(self, req_id: str) -> None:
+        """Drop any cached/in-flight lookup for ``req_id`` (e.g. on abort)."""
+        future = self.futures.pop(req_id, None)
+        if future is not None:
+            future.cancel()
+
+    def _reset(self) -> bool:
        """Trigger ``store.remove_all(force=True)`` on worker rank 0.

        Ordering assumption: caller MUST ensure no in-flight Mooncake
@@ -1582,7 +1624,11 @@ class LookupKeyClient:
        resp = self.socket.recv()
        return bytes(resp) == RESP_OK

+    def reset(self) -> bool:
+        return self.executor.submit(self._reset).result()
+
    def close(self):
+        self.executor.shutdown(wait=False, cancel_futures=True)
        self.socket.close(linger=0)


@@ -841,8 +841,106 @@ class NixlBaseConnectorWorker:
        # Forwarding a real layer name rather than a synthetic key
        self.register_kv_caches({first_layer: kv_cache})

+    def _register_packed_kv_cache(
+        self,
+        storage: torch.UntypedStorage,
+    ) -> None:
+        """Register a packed KV cache as a single NIXL region.
+
+        The packed allocation interleaves all layers per block, so each
+        block_stride-byte chunk is one logical block.  We register 1
+        NIXL region and create 1 descriptor per block.
+        """
+        self.transfer_topo = TransferTopology(
+            tp_rank=self.tp_rank,
+            tp_size=self.world_size,
+            block_size=self.block_size,
+            engine_id=self.engine_id,
+            is_mla=self.use_mla,
+            total_num_kv_heads=self.model_config.get_total_num_kv_heads(),
+            attn_backends=self.attn_backends,
+            tensor_shape=None,
+            is_mamba=self._has_mamba,
+        )
+        self.compat_hash = compute_nixl_compatibility_hash(
+            self.vllm_config,
+            self.backend_name,
+            self.transfer_topo.cross_layers_blocks,
+        )
+
+        total_size = storage.nbytes()
+        block_stride = total_size // self.num_blocks
+        base_addr = storage.data_ptr()
+        device_id = storage.device.index
+        assert device_id is not None
+
+        logger.info(
+            "Registering packed KV cache: total_size=%s, block_stride=%s, "
+            "num_blocks=%s, num_regions=1",
+            total_size,
+            block_stride,
+            self.num_blocks,
+        )
+
+        self.device_id = device_id
+        caches_data = [(base_addr, total_size, self.device_id, "")]
+
+        self.block_len_per_layer = [block_stride]
+        self.num_regions = 1
+        self.num_descs = self.num_blocks
+        self.kv_caches_base_addr[self.engine_id][self.tp_rank] = [base_addr]
+
+        descs = self.nixl_wrapper.get_reg_descs(caches_data, self.nixl_memory_type)
+        self.nixl_wrapper.register_memory(descs, backends=self.nixl_backends)
+        self._registered_descs.append(descs)
+
+        self.dst_num_blocks[self.engine_id] = self.num_blocks
+
+        self.src_xfer_handles_by_block_size[self.block_size], (self.src_blocks_data) = (
+            self.register_local_xfer_handler(self.block_size)
+        )
+
+        agent_metadata = NixlAgentMetadata(
+            engine_id=self.engine_id,
+            agent_metadata=self.nixl_wrapper.get_agent_metadata(),
+            device_id=self.device_id,
+            kv_caches_base_addr=(
+                self.kv_caches_base_addr[self.engine_id][self.tp_rank]
+            ),
+            num_blocks=self.num_blocks,
+            block_lens=self.block_len_per_layer,
+            kv_cache_layout=self.kv_cache_layout,
+            block_size=self.block_size,
+            ssm_sizes=self._mamba_ssm_size,
+            attn_backend_name=self.backend_name,
+            physical_blocks_per_logical_kv_block=(
+                self._physical_blocks_per_logical_kv_block
+            ),
+        )
+        assert self.compat_hash is not None
+        encoder = msgspec.msgpack.Encoder()
+        self.xfer_handshake_metadata = NixlHandshakePayload(
+            compatibility_hash=self.compat_hash,
+            agent_metadata_bytes=encoder.encode(agent_metadata),
+        )
+
    def register_kv_caches(self, kv_caches: dict[str, torch.Tensor]):
        """Register the KV Cache data in nixl."""
+
+        # Detect packed allocation: all tensors are strided views into the
+        # same backing storage (different data_ptr but same storage).
+        # This happens with DSv4-style contiguous per-block packing.
+        if len(kv_caches) > 1 and not self._has_mamba:
+            storage = next(iter(kv_caches.values())).untyped_storage()
+            storage_ptrs = {
+                cache.untyped_storage().data_ptr() for cache in kv_caches.values()
+            }
+            data_ptrs = {cache.data_ptr() for cache in kv_caches.values()}
+            if len(storage_ptrs) == 1 and len(data_ptrs) > 1:
+                self._register_packed_kv_cache(storage)
+                self.device_kv_caches = kv_caches
+                return
+
        self.transfer_topo = TransferTopology(
            tp_rank=self.tp_rank,
            tp_size=self.world_size,
@@ -72,18 +72,22 @@ class OffloadingConnectorWorker:
                if isinstance(layer_kv_cache_spec, AttentionSpec):
                    layer_kv_cache = kv_caches[layer_name]
                    assert isinstance(layer_kv_cache, torch.Tensor)
-                    assert layer_kv_cache.storage_offset() == 0

-                    storage = layer_kv_cache.untyped_storage()
                    page = layer_kv_cache_spec.page_size_bytes
+                    elem_size = layer_kv_cache.element_size()
+                    byte_offset = layer_kv_cache.storage_offset() * elem_size
+                    block_stride_bytes = layer_kv_cache.stride(0) * elem_size
                    tensors_per_block[layer_name] = (
                        torch.tensor(
                            [],
                            dtype=torch.int8,
                            device=layer_kv_cache.device,
-                        )
-                        .set_(storage)
-                        .view(num_blocks, page),
+                        ).set_(
+                            layer_kv_cache.untyped_storage(),
+                            byte_offset,
+                            (num_blocks, page),
+                            (block_stride_bytes, 1),
+                        ),
                    )
                    page_size_bytes[layer_name] = layer_kv_cache_spec.page_size_bytes
                    unpadded_page_size_bytes[layer_name] = (
@@ -392,6 +392,14 @@ class GroupCoordinator:

        self.rank = torch.distributed.get_rank()
        self.local_rank = local_rank
+        self.device_index: int
+        if _WORLD is not None:
+            self.device_index = _WORLD.device_index
+        else:
+            assert local_rank >= 0, (
+                "local_rank must be provided when creating the world group"
+            )
+            self.device_index = local_rank

        self_device_group = None
        self_cpu_group = None
@@ -442,11 +450,18 @@ class GroupCoordinator:
        from vllm.platforms import current_platform

        if current_platform.is_cuda_alike():
-            self.device = torch.device(f"cuda:{local_rank}")
+            visible_device_index = (
+                current_platform.logical_device_id_to_visible_device_id(
+                    self.device_index
+                )
+            )
+            self.device = torch.device(f"cuda:{visible_device_index}")
        elif current_platform.is_xpu():
-            self.device = torch.device(f"xpu:{local_rank}")
+            self.device = torch.device(f"xpu:{self.device_index}")
        elif current_platform.is_out_of_tree():
-            self.device = torch.device(f"{current_platform.device_name}:{local_rank}")
+            self.device = torch.device(
+                f"{current_platform.device_name}:{self.device_index}"
+            )
        else:
            self.device = torch.device("cpu")

@@ -1438,7 +1453,12 @@ def _init_process_group_for_split_group(
    """
    if torch.accelerator.is_available() and backend != "gloo":
        init_backend = "cpu:gloo,cuda:nccl"
-        device_id: torch.device | None = torch.device(f"cuda:{local_rank}")
+        from vllm.platforms import current_platform
+
+        visible_device_index = current_platform.logical_device_id_to_visible_device_id(
+            local_rank
+        )
+        device_id: torch.device | None = torch.device(f"cuda:{visible_device_index}")
    else:
        init_backend = "gloo"
        device_id = None
@@ -86,6 +86,15 @@ class StatelessGroupCoordinator(GroupCoordinator):

        self.rank = global_rank
        self.local_rank = local_rank
+        from vllm.distributed.parallel_state import _WORLD
+
+        if _WORLD is not None:
+            self.device_index = _WORLD.device_index
+        else:
+            assert local_rank >= 0, (
+                "local_rank must be provided when creating the world group"
+            )
+            self.device_index = local_rank

        self_device_group = None
        self_cpu_group = None
@@ -152,11 +161,18 @@ class StatelessGroupCoordinator(GroupCoordinator):
        self.tcp_store_group = self_tcp_store_group

        if current_platform.is_cuda_alike():
-            self.device = torch.device(f"cuda:{local_rank}")
+            visible_device_index = (
+                current_platform.logical_device_id_to_visible_device_id(
+                    self.device_index
+                )
+            )
+            self.device = torch.device(f"cuda:{visible_device_index}")
        elif current_platform.is_xpu():
-            self.device = torch.device(f"xpu:{local_rank}")
+            self.device = torch.device(f"xpu:{self.device_index}")
        elif current_platform.is_out_of_tree():
-            self.device = torch.device(f"{current_platform.device_name}:{local_rank}")
+            self.device = torch.device(
+                f"{current_platform.device_name}:{self.device_index}"
+            )
        else:
            self.device = torch.device("cpu")

@@ -6,6 +6,7 @@ import copy
 import dataclasses
 import functools
 import json
+import os
 import sys
 from collections.abc import Callable
 from dataclasses import MISSING, asdict, dataclass, fields, is_dataclass
@@ -465,6 +466,7 @@ class EngineArgs:
    numa_bind: bool = ParallelConfig.numa_bind
    numa_bind_nodes: list[int] | None = ParallelConfig.numa_bind_nodes
    numa_bind_cpus: list[str] | None = ParallelConfig.numa_bind_cpus
+    device_ids: list[int | str] | None = None
    tensor_parallel_size: int = ParallelConfig.tensor_parallel_size
    prefill_context_parallel_size: int = ParallelConfig.prefill_context_parallel_size
    decode_context_parallel_size: int = ParallelConfig.decode_context_parallel_size
@@ -979,6 +981,20 @@ class EngineArgs:
        parallel_group.add_argument(
            "--numa-bind-cpus", **parallel_kwargs["numa_bind_cpus"]
        )
+        parallel_group.add_argument(
+            "--device-ids",
+            type=lambda s: [
+                int(device_id) if device_id.isdigit() else device_id
+                for device_id in (part.strip() for part in s.split(","))
+            ],
+            default=None,
+            help="Comma-separated physical GPU device IDs or UUIDs to use "
+            '(e.g. --device-ids "2,3,5,7"). Avoids setting '
+            "CUDA_VISIBLE_DEVICES, preserving full GPU topology "
+            "visibility for GPU-NIC affinity and DeepGEMM. "
+            "Note: has no effect with Ray executors; use Ray "
+            "placement groups for GPU selection instead.",
+        )
        parallel_group.add_argument(
            "--tensor-parallel-size", "-tp", **parallel_kwargs["tensor_parallel_size"]
        )
@@ -1716,6 +1732,47 @@ class EngineArgs:
        )
        return SpeculativeConfig(**self.speculative_config)

+    def _resolve_device_ids(self) -> list[int] | None:
+        if not self.device_ids:
+            return None
+        if self.distributed_executor_backend == "ray":
+            logger.warning(
+                "--device-ids has no effect when using the Ray executor. "
+                "Use Ray placement groups for GPU selection instead."
+            )
+        ids = self.device_ids
+        if len(set(ids)) != len(ids):
+            raise ValueError(f"--device-ids must not contain duplicates: {ids}")
+        if all(isinstance(i, str) for i in ids):
+            return [
+                current_platform.device_control_id_to_physical_device_id(i)
+                for i in cast(list[str], ids)
+            ]
+        if any(isinstance(i, str) for i in ids):
+            raise ValueError("--device-ids must not mix integer IDs and UUIDs")
+        int_ids = cast(list[int], ids)
+        # Compose with CUDA_VISIBLE_DEVICES: if CVD is set, treat
+        # --device-ids values as indices into the CVD-visible set.
+        cvd = getattr(
+            envs,
+            current_platform.device_control_env_var,
+            os.environ.get(current_platform.device_control_env_var),
+        )
+        if cvd:
+            cvd_ids = [
+                current_platform.device_control_id_to_physical_device_id(x)
+                for x in cvd.split(",")
+            ]
+            for i in int_ids:
+                if i >= len(cvd_ids):
+                    raise ValueError(
+                        f"--device-ids index {i} is out of range for "
+                        f"{current_platform.device_control_env_var}"
+                        f"={cvd} ({len(cvd_ids)} devices visible)"
+                    )
+            return [cvd_ids[i] for i in int_ids]
+        return int_ids
+
    def create_diffusion_config(self) -> DiffusionConfig | None:
        if self.diffusion_config is None:
            return None
@@ -2029,6 +2086,7 @@ class EngineArgs:
            cp_kv_cache_interleave_size=self.cp_kv_cache_interleave_size,
            _api_process_count=self._api_process_count,
            _api_process_rank=self._api_process_rank,
+            assigned_physical_gpu_ids=self._resolve_device_ids(),
            numa_bind=self.numa_bind,
            numa_bind_nodes=self.numa_bind_nodes,
            numa_bind_cpus=self.numa_bind_cpus,
@@ -12,6 +12,7 @@ import uuid
 from collections.abc import AsyncGenerator
 from typing import TYPE_CHECKING, Any

+import jinja2
 from fastapi import Request

 from vllm.engine.protocol import EngineClient
@@ -99,6 +100,36 @@ class AnthropicServingMessages(OpenAIServingChat):
            "length": "max_tokens",
            "tool_calls": "tool_use",
        }
+        self._merge_inline_system = self._detect_merge_inline_system(chat_template)
+
+    @staticmethod
+    def _detect_merge_inline_system(chat_template: str | None) -> bool:
+        """Auto-detect whether the chat template requires system-first ordering.
+
+        Renders a [system, user, system, user] conversation against the
+        template; if it raises (e.g. Qwen's ``loop.first`` guard), the
+        model needs inline system messages merged into the leading block.
+        """
+        if not chat_template:
+            return True
+        try:
+            env = jinja2.sandbox.ImmutableSandboxedEnvironment(
+                trim_blocks=True,
+                lstrip_blocks=True,
+                extensions=[jinja2.ext.loopcontrols],
+            )
+            env.from_string(chat_template).render(
+                messages=[
+                    {"role": "system", "content": "t"},
+                    {"role": "user", "content": "t"},
+                    {"role": "system", "content": "t"},
+                    {"role": "user", "content": "t"},
+                ],
+                add_generation_prompt=False,
+            )
+            return False
+        except jinja2.TemplateError:
+            return True

    @staticmethod
    def _convert_image_source_to_url(source: dict[str, Any]) -> str:
@@ -123,13 +154,24 @@ class AnthropicServingMessages(OpenAIServingChat):

    @classmethod
    def _convert_anthropic_to_openai_request(
-        cls, anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest
+        cls,
+        anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest,
+        *,
+        merge_inline_system: bool = False,
    ) -> ChatCompletionRequest:
        """Convert Anthropic message format to OpenAI format"""
        openai_messages: list[dict[str, Any]] = []

-        cls._convert_system_message(anthropic_request, openai_messages)
-        cls._convert_messages(anthropic_request.messages, openai_messages)
+        cls._convert_system_message(
+            anthropic_request,
+            openai_messages,
+            merge_inline_system=merge_inline_system,
+        )
+        cls._convert_messages(
+            anthropic_request.messages,
+            openai_messages,
+            merge_inline_system=merge_inline_system,
+        )
        req = cls._build_base_request(anthropic_request, openai_messages)
        cls._handle_streaming_options(req, anthropic_request)
        cls._handle_output_config(req, anthropic_request)
@@ -142,6 +184,8 @@ class AnthropicServingMessages(OpenAIServingChat):
        cls,
        anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest,
        openai_messages: list[dict[str, Any]],
+        *,
+        merge_inline_system: bool = False,
    ) -> None:
        """Convert Anthropic system message to OpenAI format"""
        system_parts: list[str] = []
@@ -159,6 +203,17 @@ class AnthropicServingMessages(OpenAIServingChat):
                            continue
                        system_parts.append(block.text)

+        # When the template requires system-first ordering, extract inline
+        # system messages from the messages array and merge them into the
+        # top-level block so the template doesn't reject them.
+        if merge_inline_system:
+            for msg in anthropic_request.messages:
+                if msg.role != "system":
+                    continue
+                text = cls._extract_system_text(msg)
+                if text:
+                    system_parts.append(text)
+
        if system_parts:
            openai_messages.append({"role": "system", "content": "".join(system_parts)})

@@ -180,7 +235,11 @@ class AnthropicServingMessages(OpenAIServingChat):

    @classmethod
    def _convert_messages(
-        cls, messages: list, openai_messages: list[dict[str, Any]]
+        cls,
+        messages: list,
+        openai_messages: list[dict[str, Any]],
+        *,
+        merge_inline_system: bool = False,
    ) -> None:
        """Convert Anthropic messages to OpenAI format"""
        for msg in messages:
@@ -190,6 +249,8 @@ class AnthropicServingMessages(OpenAIServingChat):
            # doesn't strip billing headers and may produce messages with
            # no "content" key.
            if msg.role == "system":
+                if merge_inline_system:
+                    continue  # already merged into top-level by _convert_system_message
                text = cls._extract_system_text(msg)
                if text:
                    openai_messages.append({"role": "system", "content": text})
@@ -497,7 +558,10 @@ class AnthropicServingMessages(OpenAIServingChat):
        """
        if logger.isEnabledFor(logging.DEBUG):
            logger.debug("Received messages request %s", request.model_dump_json())
-        chat_req = self._convert_anthropic_to_openai_request(request)
+        chat_req = self._convert_anthropic_to_openai_request(
+            request,
+            merge_inline_system=self._merge_inline_system,
+        )
        if logger.isEnabledFor(logging.DEBUG):
            logger.debug("Convert to OpenAI request %s", chat_req.model_dump_json())
        generator = await self.create_chat_completion(chat_req, raw_request)
@@ -905,7 +969,10 @@ class AnthropicServingMessages(OpenAIServingChat):
        raw_request: Request | None = None,
    ) -> AnthropicCountTokensResponse | ErrorResponse:
        """Implements Anthropic's messages.count_tokens endpoint."""
-        chat_req = self._convert_anthropic_to_openai_request(request)
+        chat_req = self._convert_anthropic_to_openai_request(
+            request,
+            merge_inline_system=self._merge_inline_system,
+        )
        result = await self.render_chat_request(chat_req)
        if isinstance(result, ErrorResponse):
            return result
@@ -898,12 +898,6 @@ class LLM(BeamSearchOfflineMixin, PoolingOfflineMixin, OfflineInferenceMixin):
    def finish_weight_update(self) -> None:
        """Finish the current weight update."""
        self.llm_engine.collective_rpc("finish_weight_update")
-        # Invalidate cached state computed with the old weights so it isn't
-        # reused for subsequent requests:
-        # - prefix cache: KV blocks computed with the old weights
-        # - encoder cache: multimodal embeddings keyed only by mm_hash
-        self.llm_engine.reset_prefix_cache()
-        self.llm_engine.reset_encoder_cache()

    def __repr__(self) -> str:
        """Return a transformers-style hierarchical view of the model."""
--- a/Show More
+++ b/Show More