low latency combine

Signed-off-by: Zhuohan Li <zhuohan123@gmail.com>
low latency dispatch
2025-10-20 22:00:01 -07:00 · 2025-10-20 21:49:09 -07:00 · 2025-10-20 18:08:57 -07:00 · 2025-10-20 17:45:31 -07:00 · 2025-10-19 23:31:03 -07:00 · 2025-10-19 23:30:28 -07:00
1428 changed files with 71626 additions and 19198 deletions
@@ -454,11 +454,6 @@ main() {
  fi
  check_hf_token

-  # Set to v1 to run v1 benchmark
-  if [[ "${ENGINE_VERSION:-v0}" == "v1" ]]; then
-    export VLLM_USE_V1=1
-  fi
-
  # dependencies
  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
  (which jq) || (apt-get update && apt-get -y install jq)
@@ -48,7 +48,7 @@ steps:
    agents:
      queue: cpu_queue_postmerge
    commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0+PTX' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
      - "mkdir artifacts"
      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
      - "bash .buildkite/scripts/upload-wheels.sh"
@@ -25,25 +25,28 @@ function cpu_tests() {

  # offline inference
  podman exec -it "$container_id" bash -c "
-    set -e
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
+    set -xve
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m" >> $HOME/test_basic.log

  # Run basic model test
  podman exec -it "$container_id" bash -c "
-    set -e
+    set -evx
    pip install pytest pytest-asyncio einops peft Pillow soundfile transformers_stream_generator matplotlib
    pip install sentence-transformers datamodel_code_generator
-    pytest -v -s tests/models/language/generation/test_bart.py -m cpu_model
+
+    # Note: disable Bart until supports V1
+    # pytest -v -s tests/models/language/generation/test_bart.py -m cpu_model
    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-openai-community/gpt2]
    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-facebook/opt-125m]
    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-google/gemma-1.1-2b-it]
    pytest -v -s tests/models/language/pooling/test_classification.py::test_models[float-jason9693/Qwen2.5-1.5B-apeach]
-    pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model"
+    # TODO: Below test case tests/models/language/pooling/test_embedding.py::test_models[True-ssmits/Qwen2-7B-Instruct-embed-base] fails on ppc64le. Disabling it for time being.
+    # pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model" >> $HOME/test_rest.log
 }

 # All of CPU tests are expected to be finished less than 40 mins.

 export container_id
 export -f cpu_tests
-timeout 40m bash -c cpu_tests
+timeout 120m bash -c cpu_tests

@@ -64,10 +64,9 @@ python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git
    && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
    && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
 echo "--- Python dependencies installed ---"
-export VLLM_USE_V1=1
+
 export VLLM_XLA_CHECK_RECOMPILATION=1
 export VLLM_XLA_CACHE_PATH=
-echo "Using VLLM V1"

 echo "--- Hardware Information ---"
 # tpu-info
@@ -64,10 +64,9 @@ python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git
    && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
    && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
 echo "--- Python dependencies installed ---"
-export VLLM_USE_V1=1
+
 export VLLM_XLA_CHECK_RECOMPILATION=1
 export VLLM_XLA_CACHE_PATH=
-echo "Using VLLM V1"

 echo "--- Hardware Information ---"
 # tpu-info
@@ -44,6 +44,5 @@ docker run \
    pytest -v -s v1/structured_output
    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py
    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_shared_storage_connector.py
-    pytest -v -s v1/test_metrics
    pytest -v -s v1/test_serial_utils.py
 '
@@ -9,6 +9,6 @@ MAX_NUM_BATCHED_TOKENS=1024
 TENSOR_PARALLEL_SIZE=1
 MAX_MODEL_LEN=2048
 DOWNLOAD_DIR=/mnt/disks/persist
-EXPECTED_THROUGHPUT=10.0
+EXPECTED_THROUGHPUT=8.7
 INPUT_LEN=1800
 OUTPUT_LEN=128
@@ -42,7 +42,7 @@ echo "lanching vllm..."
 echo "logging to $VLLM_LOG"
 echo

-VLLM_USE_V1=1 vllm serve $MODEL \
+vllm serve $MODEL \
 --seed 42 \
 --max-num-seqs $MAX_NUM_SEQS \
 --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
@@ -296,6 +296,7 @@ steps:
    - tests/v1
  commands:
    # split the test to avoid interference
+    - pytest -v -s -m 'not cpu_test' v1/core
    - pytest -v -s v1/executor
    - pytest -v -s v1/kv_offload
    - pytest -v -s v1/sample
@@ -317,7 +318,7 @@ steps:
  no_gpu: true
  commands:
    # split the test to avoid interference
-    - pytest -v -s v1/core
+    - pytest -v -s -m 'cpu_test' v1/core
    - pytest -v -s v1/structured_output
    - pytest -v -s v1/test_serial_utils.py
    - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
@@ -399,11 +400,10 @@ steps:
    - pytest -v -s compile/test_fusion_attn.py
    - pytest -v -s compile/test_functionalization.py
    - pytest -v -s compile/test_silu_mul_quant_fusion.py
-    - pytest -v -s compile/test_sequence_parallelism.py
-    - pytest -v -s compile/test_async_tp.py
    - pytest -v -s compile/test_fusion_all_reduce.py
    - pytest -v -s compile/test_decorator.py
    - pytest -v -s compile/test_noop_elimination.py
+    - pytest -v -s compile/test_aot_compile.py

 - label: PyTorch Fullgraph Smoke Test # 15min
  timeout_in_minutes: 30
@@ -432,8 +432,9 @@ steps:
  source_file_dependencies:
  - csrc/
  - tests/kernels/core
+  - tests/kernels/test_top_k_per_row.py
  commands:
-    - pytest -v -s kernels/core
+    - pytest -v -s kernels/core kernels/test_top_k_per_row.py

 - label: Kernels Attention Test %N # 23min
  timeout_in_minutes: 35
@@ -828,12 +829,14 @@ steps:
    - pytest -v -s tests/kernels/quantization/test_flashinfer_scaled_mm.py
    - pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py
    - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
-    - pytest -v -s tests/kernels/moe/test_mxfp4_moe.py
+    - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
    # Fusion
    - pytest -v -s tests/compile/test_fusion_all_reduce.py
    - pytest -v -s tests/compile/test_fusion_attn.py::test_attention_quant_pattern
    - pytest -v -s tests/kernels/moe/test_flashinfer.py
    - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
+    - pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py
+    - pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py

 - label: Blackwell GPT-OSS Eval
  timeout_in_minutes: 60
@@ -867,7 +870,7 @@ steps:
    - pytest -s -v tests/quantization/test_blackwell_moe.py

 - label: Blackwell LM Eval Small Models
-  timeout_in_minutes: 75
+  timeout_in_minutes: 120
  gpu: b200
  optional: true # run on nightlies
  source_file_dependencies:
@@ -1092,6 +1095,8 @@ steps:
  working_dir: "/vllm-workspace/"
  num_gpus: 2
  commands:
+    - pytest -v -s tests/compile/test_async_tp.py
+    - pytest -v -s tests/compile/test_sequence_parallelism.py
    - pytest -v -s tests/distributed/test_context_parallel.py
    - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048

@@ -121,3 +121,11 @@ mkdocs.yaml @hmellor

 # KVConnector installation files
 /requirements/kv_connectors.txt @NickLucche
+
+# Pooling models
+/examples/*/pooling/ @noooop
+/tests/models/*/pooling* @noooop
+/tests/entrypoints/pooling @noooop
+/vllm/config/pooler.py @noooop
+/vllm/pooling_params.py @noooop
+/vllm/model_executor/layers/pooler.py @noooop
@@ -11,6 +11,8 @@ pull_request_rules:
    label:
      add:
        - documentation
+    comment:
+      message: "Documentation preview: https://vllm--{{number}}.org.readthedocs.build/en/{{number}}/"

 - name: label-ci-build
  description: Automatically apply ci/build label
@@ -7,17 +7,17 @@ default_stages:
 exclude: 'vllm/third_party/.*'
 repos:
 - repo: https://github.com/astral-sh/ruff-pre-commit
-  rev: v0.13.3
+  rev: v0.14.0
  hooks:
  - id: ruff-check
    args: [--output-format, github, --fix]
  - id: ruff-format
 - repo: https://github.com/crate-ci/typos
-  rev: v1.35.5
+  rev: v1.38.1
  hooks:
  - id: typos
 - repo: https://github.com/pre-commit/mirrors-clang-format
-  rev: v20.1.3
+  rev: v21.1.2
  hooks:
  - id: clang-format
    exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
@@ -34,7 +34,7 @@ repos:
  hooks:
  - id: actionlint
 - repo: https://github.com/astral-sh/uv-pre-commit
-  rev: 0.6.17
+  rev: 0.9.1
  hooks:
    - id: pip-compile
      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128, --python-platform, x86_64-manylinux_2_28]
@@ -55,11 +55,6 @@ repos:
      types_or: [python, pyi]
      require_serial: true
      additional_dependencies: [mypy==1.11.1, regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
-  - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
-    name: Run mypy for Python 3.9
-    entry: python tools/pre_commit/mypy.py 1 "3.9"
-    <<: *mypy_common
-    stages: [manual] # Only run in CI
  - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
    name: Run mypy for Python 3.10
    entry: python tools/pre_commit/mypy.py 1 "3.10"
@@ -75,6 +70,11 @@ repos:
    entry: python tools/pre_commit/mypy.py 1 "3.12"
    <<: *mypy_common
    stages: [manual] # Only run in CI
+  - id: mypy-3.13 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.13
+    entry: python tools/pre_commit/mypy.py 1 "3.13"
+    <<: *mypy_common
+    stages: [manual] # Only run in CI
  - id: shellcheck
    name: Lint shell scripts
    entry: tools/shellcheck.sh
@@ -34,7 +34,7 @@ install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
 #
-set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12" "3.13")
+set(PYTHON_SUPPORTED_VERSIONS "3.10" "3.11" "3.12" "3.13")

 # Supported AMD GPU architectures.
 set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151")
@@ -269,8 +269,8 @@ set(VLLM_EXT_SRC
  "csrc/sampler.cu"
  "csrc/cuda_view.cu"
  "csrc/quantization/gptq/q_gemm.cu"
-  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
-  "csrc/quantization/fp8/common.cu"
+  "csrc/quantization/w8a8/int8/scaled_quant.cu"
+  "csrc/quantization/w8a8/fp8/common.cu"
  "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
  "csrc/quantization/gguf/gguf_kernel.cu"
  "csrc/quantization/activation_kernels.cu"
@@ -314,12 +314,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  list(APPEND VLLM_EXT_SRC
    "csrc/quantization/awq/gemm_kernels.cu"
    "csrc/permute_cols.cu"
-    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
+    "csrc/quantization/w8a8/cutlass/scaled_mm_entry.cu"
    "csrc/quantization/fp4/nvfp4_quant_entry.cu"
    "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
    "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
    "csrc/cutlass_extensions/common.cpp"
-    "csrc/quantization/fp8/per_token_group_quant.cu")
+    "csrc/quantization/w8a8/fp8/per_token_group_quant.cu"
+    "csrc/quantization/w8a8/int8/per_token_group_quant.cu")

  set_gencode_flags_for_srcs(
    SRCS "${VLLM_EXT_SRC}"
@@ -423,11 +424,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
    set(SRCS
-       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
+       "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm90.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_fp8.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_int8.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_azp_sm90_int8.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm90_fp8.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -458,9 +459,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
    set(SRCS
-      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm120_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm120.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm120_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu"
    )
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
@@ -492,9 +493,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
    set(SRCS
-      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm100.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm100_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm100_fp8.cu"
    )
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
@@ -525,7 +526,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  # subtract out the archs that are already built for 3x
  list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
  if (SCALED_MM_2X_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/scaled_mm_c2x.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_2X_ARCHS}")
@@ -648,7 +649,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
  # if it's possible to compile MoE kernels that use its output.
  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm90.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm90.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -672,7 +673,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm100.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -697,7 +698,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/moe_data.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
@@ -720,7 +721,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/blockwise_scaled_group_mm_sm100.cu")
    set_gencode_flags_for_srcs(
      SRCS "${SRCS}"
      CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -1006,6 +1007,7 @@ endif()
 # For CUDA we also build and ship some external projects.
 if (VLLM_GPU_LANG STREQUAL "CUDA")
    include(cmake/external_projects/flashmla.cmake)
+    include(cmake/external_projects/qutlass.cmake)

    # vllm-flash-attn should be last as it overwrites some CMake functions
    include(cmake/external_projects/vllm_flash_attn.cmake)
@@ -149,6 +149,7 @@ Compute Resources:
 - Trainy
 - UC Berkeley
 - UC San Diego
+- Volcengine

 Slack Sponsor: Anyscale

@@ -74,7 +74,7 @@ start_server() {
    local vllm_log=$4
    local profile_dir=$5

-    pkill -if vllm
+    pkill -if "vllm serve" || true

    # Define the common arguments as a bash array.
    # Each argument and its value are separate elements.
@@ -96,11 +96,11 @@ start_server() {
    # This correctly passes each element as a separate argument.
    if [[ -n "$profile_dir" ]]; then
        # Start server with profiling enabled
-        VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir \
+        VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir \
            vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
    else
        # Start server without profiling
-        VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 \
+        VLLM_SERVER_DEV_MODE=1 \
            vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
    fi
    local server_pid=$!
@@ -139,7 +139,7 @@ run_benchmark() {
    echo "vllm_log: $vllm_log"
    echo
    rm -f $vllm_log
-    pkill -if vllm
+    pkill -if "vllm serve" || true

    echo "starting server..."
    # Call start_server without a profile_dir to avoid profiling overhead
@@ -232,7 +232,7 @@ run_benchmark() {

    echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"

-    pkill -if vllm
+    pkill -if "vllm serve" || true
    sleep 10
    echo "===================="
    return 0
@@ -308,6 +308,6 @@ if (( $(echo "$best_throughput > 0" | bc -l) )); then
 else
    echo "No configuration met the latency requirements. Skipping final profiling run."
 fi
-pkill -if vllm
+pkill -if "vllm serve" || true
 echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH"
 echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT"
@@ -8,7 +8,6 @@ import sys
 import time
 import traceback
 from dataclasses import dataclass, field
-from typing import Optional, Union

 import aiohttp
 import huggingface_hub.constants
@@ -28,13 +27,13 @@ class RequestFuncInput:
    prompt_len: int
    output_len: int
    model: str
-    model_name: Optional[str] = None
-    logprobs: Optional[int] = None
-    extra_body: Optional[dict] = None
-    multi_modal_content: Optional[dict | list[dict]] = None
+    model_name: str | None = None
+    logprobs: int | None = None
+    extra_body: dict | None = None
+    multi_modal_content: dict | list[dict] | None = None
    ignore_eos: bool = False
-    language: Optional[str] = None
-    request_id: Optional[str] = None
+    language: str | None = None
+    request_id: str | None = None


@dataclass
@@ -52,7 +51,7 @@ class RequestFuncOutput:

 async def async_request_tgi(
    request_func_input: RequestFuncInput,
-    pbar: Optional[tqdm] = None,
+    pbar: tqdm | None = None,
 ) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    assert api_url.endswith("generate_stream")
@@ -133,7 +132,7 @@ async def async_request_tgi(

 async def async_request_trt_llm(
    request_func_input: RequestFuncInput,
-    pbar: Optional[tqdm] = None,
+    pbar: tqdm | None = None,
 ) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    assert api_url.endswith("generate_stream")
@@ -204,7 +203,7 @@ async def async_request_trt_llm(

 async def async_request_deepspeed_mii(
    request_func_input: RequestFuncInput,
-    pbar: Optional[tqdm] = None,
+    pbar: tqdm | None = None,
 ) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    assert api_url.endswith(("completions", "profile")), (
@@ -267,7 +266,7 @@ async def async_request_deepspeed_mii(

 async def async_request_openai_completions(
    request_func_input: RequestFuncInput,
-    pbar: Optional[tqdm] = None,
+    pbar: tqdm | None = None,
 ) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    assert api_url.endswith(("completions", "profile")), (
@@ -367,7 +366,7 @@ async def async_request_openai_completions(

 async def async_request_openai_chat_completions(
    request_func_input: RequestFuncInput,
-    pbar: Optional[tqdm] = None,
+    pbar: tqdm | None = None,
 ) -> RequestFuncOutput:
    api_url = request_func_input.api_url
    assert api_url.endswith(("chat/completions", "profile")), (
@@ -476,7 +475,7 @@ async def async_request_openai_chat_completions(

 async def async_request_openai_audio(
    request_func_input: RequestFuncInput,
-    pbar: Optional[tqdm] = None,
+    pbar: tqdm | None = None,
 ) -> RequestFuncOutput:
    # Lazy import without PlaceholderModule to avoid vllm dep.
    import soundfile
@@ -610,7 +609,7 @@ def get_tokenizer(
    tokenizer_mode: str = "auto",
    trust_remote_code: bool = False,
    **kwargs,
-) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
+) -> PreTrainedTokenizer | PreTrainedTokenizerFast:
    if pretrained_model_name_or_path is not None and not os.path.exists(
        pretrained_model_name_or_path
    ):
@@ -32,7 +32,6 @@ import dataclasses
 import json
 import random
 import time
-from typing import Optional

 from transformers import PreTrainedTokenizerBase

@@ -80,7 +79,7 @@ def sample_requests_from_dataset(
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    input_length_range: tuple[int, int],
-    fixed_output_len: Optional[int],
+    fixed_output_len: int | None,
 ) -> list[Request]:
    if fixed_output_len is not None and fixed_output_len < 4:
        raise ValueError("output_len too small")
@@ -128,7 +127,7 @@ def sample_requests_from_random(
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
    input_length_range: tuple[int, int],
-    fixed_output_len: Optional[int],
+    fixed_output_len: int | None,
    prefix_len: int,
 ) -> list[Request]:
    requests = []
@@ -7,7 +7,6 @@ import dataclasses
 import json
 import random
 import time
-from typing import Optional

 from transformers import AutoTokenizer, PreTrainedTokenizerBase

@@ -24,7 +23,7 @@ def sample_requests(
    dataset_path: str,
    num_requests: int,
    tokenizer: PreTrainedTokenizerBase,
-    fixed_output_len: Optional[int],
+    fixed_output_len: int | None,
 ) -> list[tuple[str, int, int, int]]:
    if fixed_output_len is not None and fixed_output_len < 4:
        raise ValueError("output_len too small")
@@ -32,7 +32,6 @@ import uuid
 import warnings
 from collections.abc import AsyncGenerator
 from dataclasses import dataclass
-from typing import Optional

 import datasets
 import numpy as np
@@ -316,7 +315,7 @@ def calculate_metrics(
    tokenizer: PreTrainedTokenizerBase,
    selected_percentile_metrics: list[str],
    selected_percentiles: list[float],
-    goodput_config_dict: Optional[dict[str, float]] = None,
+    goodput_config_dict: dict[str, float] | None = None,
 ) -> tuple[BenchmarkMetrics, list[int]]:
    actual_output_lens: list[int] = []
    total_input = 0
@@ -436,9 +435,9 @@ async def benchmark(
    selected_percentile_metrics: list[str],
    selected_percentiles: list[str],
    ignore_eos: bool,
-    max_concurrency: Optional[int],
+    max_concurrency: int | None,
    structured_output_ratio: float,
-    goodput_config_dict: Optional[dict[str, float]] = None,
+    goodput_config_dict: dict[str, float] | None = None,
 ):
    if backend in ASYNC_REQUEST_FUNCS:
        request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -6,7 +6,7 @@ import math
 import os
 import time
 from types import TracebackType
-from typing import Any, Optional, Union
+from typing import Any


 def convert_to_pytorch_benchmark_format(
@@ -92,7 +92,7 @@ class TimeCollector:
    def __init__(self, scale: int) -> None:
        self.cnt: int = 0
        self._sum: int = 0
-        self._max: Optional[int] = None
+        self._max: int | None = None
        self.scale = scale
        self.start_time: int = time.monotonic_ns()

@@ -104,13 +104,13 @@ class TimeCollector:
        else:
            self._max = max(self._max, v)

-    def avg(self) -> Union[float, str]:
+    def avg(self) -> float | str:
        return self._sum * 1.0 / self.cnt / self.scale if self.cnt > 0 else "N/A"

-    def max(self) -> Union[float, str]:
+    def max(self) -> float | str:
        return self._max / self.scale if self._max else "N/A"

-    def dump_avg_max(self) -> list[Union[float, str]]:
+    def dump_avg_max(self) -> list[float | str]:
        return [self.avg(), self.max()]

    def __enter__(self) -> None:
@@ -118,8 +118,8 @@ class TimeCollector:

    def __exit__(
        self,
-        exc_type: Optional[type[BaseException]],
-        exc_value: Optional[BaseException],
-        exc_traceback: Optional[TracebackType],
+        exc_type: type[BaseException] | None,
+        exc_value: BaseException | None,
+        exc_traceback: TracebackType | None,
    ) -> None:
        self.collect(time.monotonic_ns() - self.start_time)
@@ -6,8 +6,7 @@ import copy
 import itertools
 import pickle as pkl
 import time
-from collections.abc import Iterable
-from typing import Callable
+from collections.abc import Callable, Iterable

 import torch
 import torch.utils.benchmark as TBenchmark
@@ -6,8 +6,7 @@ import copy
 import itertools
 import pickle as pkl
 import time
-from collections.abc import Iterable
-from typing import Callable, Optional
+from collections.abc import Callable, Iterable

 import torch
 import torch.utils.benchmark as TBenchmark
@@ -53,7 +52,7 @@ def bench_int8(
    n: int,
    label: str,
    sub_label: str,
-    bench_kernels: Optional[list[str]] = None,
+    bench_kernels: list[str] | None = None,
 ) -> Iterable[TMeasurement]:
    """Benchmark INT8-based kernels."""
    assert dtype == torch.int8
@@ -108,7 +107,7 @@ def bench_fp8(
    n: int,
    label: str,
    sub_label: str,
-    bench_kernels: Optional[list[str]] = None,
+    bench_kernels: list[str] | None = None,
 ) -> Iterable[TMeasurement]:
    """Benchmark FP8-based kernels."""
    assert dtype == torch.float8_e4m3fn
@@ -183,7 +182,7 @@ def bench(
    n: int,
    label: str,
    sub_label: str,
-    bench_kernels: Optional[list[str]] = None,
+    bench_kernels: list[str] | None = None,
 ) -> Iterable[TMeasurement]:
    if dtype == torch.int8:
        return bench_int8(dtype, m, k, n, label, sub_label, bench_kernels)
@@ -201,7 +200,7 @@ def print_timers(timers: Iterable[TMeasurement]):
 def run(
    dtype: torch.dtype,
    MKNs: Iterable[tuple[int, int, int]],
-    bench_kernels: Optional[list[str]] = None,
+    bench_kernels: list[str] | None = None,
 ) -> Iterable[TMeasurement]:
    results = []
    for m, k, n in MKNs:
@@ -3,10 +3,9 @@

 import pickle as pkl
 import time
-from collections.abc import Iterable
+from collections.abc import Callable, Iterable
 from dataclasses import dataclass
 from itertools import product
-from typing import Callable, Optional

 import torch
 import torch.utils.benchmark as TBenchmark
@@ -51,7 +50,7 @@ def get_bench_params() -> list[bench_params_t]:
 def unfused_int8_impl(
    rms_norm_layer: RMSNorm,
    x: torch.Tensor,
-    residual: Optional[torch.Tensor],
+    residual: torch.Tensor | None,
    quant_dtype: torch.dtype,
 ):
    # Norm
@@ -68,7 +67,7 @@ def unfused_int8_impl(
 def unfused_fp8_impl(
    rms_norm_layer: RMSNorm,
    x: torch.Tensor,
-    residual: Optional[torch.Tensor],
+    residual: torch.Tensor | None,
    quant_dtype: torch.dtype,
 ):
    # Norm
@@ -85,7 +84,7 @@ def unfused_fp8_impl(
 def fused_impl(
    rms_norm_layer: RMSNorm,  # this stores the weights
    x: torch.Tensor,
-    residual: Optional[torch.Tensor],
+    residual: torch.Tensor | None,
    quant_dtype: torch.dtype,
 ):
    out, _ = ops.rms_norm_dynamic_per_token_quant(
@@ -0,0 +1,191 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Copyright (C) 2025 Roberto L. Castro (Roberto.LopezCastro@ist.ac.at).
+# All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#       http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import argparse
+import copy
+import itertools
+
+import torch
+from compressed_tensors.transform.utils.hadamard import deterministic_hadamard_matrix
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm._custom_ops import fusedQuantizeMx, matmul_mxf4_bf16_tn
+from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
+from vllm.triton_utils import triton
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "mxfp4": dict(no_a_quant=False, enabled=True),
+    "mxfp4-noquant": dict(no_a_quant=True, enabled=True),
+}
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
+
+
+def get_hadamard_matrix(group_size: int, dtype: torch.dtype, device: torch.device):
+    return (
+        deterministic_hadamard_matrix(group_size, dtype=dtype, device=device)
+        * group_size**-0.5
+    )
+
+
+def _quant_weight_mxfp4(
+    b: torch.Tensor, forward_hadamard_matrix: torch.Tensor, device: str
+):
+    weight_hf_e2m1, weight_hf_e8m0 = fusedQuantizeMx(
+        b, forward_hadamard_matrix, method="abs_max"
+    )
+    weight_hf_scale_block = to_blocked(weight_hf_e8m0, backend="triton")
+    return weight_hf_e2m1, weight_hf_scale_block
+
+
+def build_mxfp4_runner(cfg, a, b, forward_hadamard_matrix, dtype, device):
+    weight_hf_e2m1, weight_hf_scale_block = _quant_weight_mxfp4(
+        b, forward_hadamard_matrix, device
+    )
+    alpha = torch.tensor([1.0], device="cuda")
+
+    if cfg["no_a_quant"]:
+        # Pre-quantize activation
+        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeMx(
+            a, forward_hadamard_matrix, method="abs_max"
+        )
+        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton")
+
+        def run():
+            return matmul_mxf4_bf16_tn(
+                input_hf_e2m1,
+                weight_hf_e2m1,
+                input_hf_scale_block,
+                weight_hf_scale_block,
+                alpha,
+            )
+
+        return run
+
+    # Quantize activation on-the-fly
+    def run():
+        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeMx(
+            a, forward_hadamard_matrix, method="abs_max"
+        )
+        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton")
+        return matmul_mxf4_bf16_tn(
+            input_hf_e2m1,
+            weight_hf_e2m1,
+            input_hf_scale_block,
+            weight_hf_scale_block,
+            alpha,
+        )
+
+    return run
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[
+            1,
+            4,
+            8,
+            16,
+            32,
+            64,
+            128,
+            256,
+            512,
+            1024,
+            2048,
+            4096,
+            8192,
+            16384,
+            24576,
+            32768,
+        ],
+        x_log=False,
+        line_arg="provider",
+        line_vals=_enabled,
+        line_names=_enabled,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs MXFP4 GEMMs",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K, had_size):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+
+    a = torch.randn((M, K), device=device, dtype=dtype)
+    b = torch.randn((N, K), device=device, dtype=dtype)
+    forward_hadamard_matrix = get_hadamard_matrix(had_size, dtype, device)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), rep=200, quantiles=quantiles
+        )
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_mxfp4_runner(
+            cfg, a, b, forward_hadamard_matrix, dtype, device
+        )
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: run_quant(), rep=200, quantiles=quantiles
+        )
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    out = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            out.append(KN)
+    return out
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.3-70B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        for had_size in [32, 64, 128]:
+            print(f"{model}, N={N} K={K}, HAD={had_size}, BF16 vs MXFP4 GEMMs TFLOP/s:")
+            benchmark.run(
+                print_data=True,
+                show_plots=True,
+                save_path=f"bench_mxfp4_res_n{N}_k{K}",
+                N=N,
+                K=K,
+                had_size=had_size,
+            )
+
+    print("Benchmark finished!")
@@ -0,0 +1,207 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Copyright (C) 2025 Roberto L. Castro (Roberto.LopezCastro@ist.ac.at).
+# All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#       http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import argparse
+import copy
+import itertools
+
+import torch
+from compressed_tensors.transform.utils.hadamard import deterministic_hadamard_matrix
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops  # use existing nvfp4 gemm in vllm
+from vllm._custom_ops import fusedQuantizeNv
+from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
+from vllm.triton_utils import triton
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "nvfp4": dict(no_a_quant=False, enabled=True),
+    "nvfp4-noquant": dict(no_a_quant=True, enabled=True),
+}
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
+
+
+def get_hadamard_matrix(group_size: int, dtype: torch.dtype, device: torch.device):
+    return (
+        deterministic_hadamard_matrix(group_size, dtype=dtype, device=device)
+        * group_size**-0.5
+    )
+
+
+def _quant_weight_nvfp4(
+    b: torch.Tensor,
+    forward_hadamard_matrix: torch.Tensor,
+    global_scale: torch.Tensor,
+    device: str,
+    M: int,
+    N: int,
+    K: int,
+):
+    weight_hf_e2m1, weight_hf_e8m0 = fusedQuantizeNv(
+        b, forward_hadamard_matrix, global_scale
+    )
+    weight_hf_scale_block = to_blocked(weight_hf_e8m0, backend="triton").view(
+        -1, K // 16
+    )
+    return weight_hf_e2m1, weight_hf_scale_block
+
+
+def build_nvfp4_runner(cfg, a, b, forward_hadamard_matrix, dtype, device, M, N, K):
+    alpha = torch.tensor([1.0], device="cuda")
+    global_scale = torch.tensor([1.0], device="cuda")
+    weight_hf_e2m1, weight_hf_scale_block = _quant_weight_nvfp4(
+        b, forward_hadamard_matrix, global_scale, device, M, N, K
+    )
+
+    if cfg["no_a_quant"]:
+        # Pre-quantize activation
+        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeNv(
+            a, forward_hadamard_matrix, global_scale
+        )
+        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton").view(
+            -1, K // 16
+        )
+
+        def run():
+            return ops.cutlass_scaled_fp4_mm(
+                input_hf_e2m1,
+                weight_hf_e2m1,
+                input_hf_scale_block,
+                weight_hf_scale_block,
+                alpha,
+                torch.bfloat16,
+            )
+
+        return run
+
+    # Quantize activation on-the-fly
+    def run():
+        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeNv(
+            a, forward_hadamard_matrix, global_scale
+        )
+        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton").view(
+            -1, K // 16
+        )
+        return ops.cutlass_scaled_fp4_mm(
+            input_hf_e2m1,
+            weight_hf_e2m1,
+            input_hf_scale_block,
+            weight_hf_scale_block,
+            alpha,
+            torch.bfloat16,
+        )
+
+    return run
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[
+            1,
+            4,
+            8,
+            16,
+            32,
+            64,
+            128,
+            256,
+            512,
+            1024,
+            2048,
+            4096,
+            8192,
+            16384,
+            24576,
+            32768,
+        ],
+        x_log=False,
+        line_arg="provider",
+        line_vals=_enabled,
+        line_names=_enabled,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs NVFP4 GEMMs",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K, had_size):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+
+    a = torch.randn((M, K), device=device, dtype=dtype)
+    b = torch.randn((N, K), device=device, dtype=dtype)
+    forward_hadamard_matrix = get_hadamard_matrix(had_size, dtype, device)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), rep=200, quantiles=quantiles
+        )
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_nvfp4_runner(
+            cfg, a, b, forward_hadamard_matrix, dtype, device, M, N, K
+        )
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: run_quant(), rep=200, quantiles=quantiles
+        )
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    out = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            out.append(KN)
+    return out
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.3-70B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        for had_size in [16, 32, 64, 128]:
+            print(f"{model}, N={N} K={K}, HAD={had_size}, BF16 vs NVFP4 GEMMs TFLOP/s:")
+            benchmark.run(
+                print_data=True,
+                show_plots=True,
+                save_path=f"bench_nvfp4_res_n{N}_k{K}",
+                N=N,
+                K=K,
+                had_size=had_size,
+            )
+
+    print("Benchmark finished!")
@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import itertools
-from typing import Callable
+from collections.abc import Callable
 from unittest.mock import patch

 import pandas as pd
@@ -22,8 +22,8 @@ Example:
 import json
 import os
 import time
+from collections.abc import Callable
 from contextlib import nullcontext
-from typing import Callable, Optional

 import torch
 import torch.distributed as dist
@@ -264,12 +264,12 @@ class CommunicatorBenchmark:
    def benchmark_allreduce_single(
        self,
        sequence_length: int,
-        allreduce_fn: Callable[[torch.Tensor], Optional[torch.Tensor]],
+        allreduce_fn: Callable[[torch.Tensor], torch.Tensor | None],
        should_use_fn: Callable[[torch.Tensor], bool],
        context,
        num_warmup: int,
        num_trials: int,
-    ) -> Optional[float]:
+    ) -> float | None:
        """Benchmark method with CUDA graph optimization."""
        try:
            # Create test tensor (2D: sequence_length x hidden_size)
@@ -6,11 +6,12 @@ import copy
 import json
 import pickle
 import time
+from collections.abc import Callable
 from dataclasses import dataclass
 from enum import Enum, auto
 from itertools import product
 from pathlib import Path
-from typing import Any, Callable, Optional
+from typing import Any

 import torch
 import torch.utils.benchmark as TBenchmark
@@ -158,7 +159,7 @@ def ref_group_gemm(
    seq_lens_cpu: torch.Tensor,
    prompt_lora_mapping_cpu: torch.Tensor,
    scaling: float,
-    add_inputs: Optional[bool],
+    add_inputs: bool | None,
 ):
    """
    Torch group gemm reference implementation to test correctness of
@@ -316,8 +317,8 @@ class BenchmarkContext:
    lora_rank: int
    sort_by_lora_id: bool
    dtype: torch.dtype
-    seq_length: Optional[int] = None
-    num_slices: Optional[int] = None  # num_slices for slice based ops
+    seq_length: int | None = None
+    num_slices: int | None = None  # num_slices for slice based ops

    def with_seq_length(self, seq_length: int) -> "BenchmarkContext":
        ctx = copy.copy(self)
@@ -561,7 +562,7 @@ class BenchmarkTensors:
        }

    def bench_fn_kwargs(
-        self, op_type: OpType, add_inputs: Optional[bool] = None
+        self, op_type: OpType, add_inputs: bool | None = None
    ) -> dict[str, Any]:
        if op_type.is_shrink_fn():
            assert add_inputs is None
@@ -575,7 +576,7 @@ class BenchmarkTensors:
        raise ValueError(f"Unrecognized optype {self}")

    def test_correctness(
-        self, op_type: OpType, expand_fn_add_inputs: Optional[bool]
+        self, op_type: OpType, expand_fn_add_inputs: bool | None
    ) -> bool:
        """
        Test correctness of op_type implementation against a grouped gemm
@@ -611,8 +612,8 @@ def bench_optype(
    ctx: BenchmarkContext,
    arg_pool_size: int,
    op_type: OpType,
-    cuda_graph_nops: Optional[int] = None,
-    expand_fn_add_inputs: Optional[bool] = None,
+    cuda_graph_nops: int | None = None,
+    expand_fn_add_inputs: bool | None = None,
    test_correctness: bool = False,
 ) -> TMeasurement:
    assert arg_pool_size >= 1
@@ -679,7 +680,7 @@ def bench_torch_mm(
    ctx: BenchmarkContext,
    arg_pool_size: int,
    op_type: OpType,
-    cuda_graph_nops: Optional[int] = None,
+    cuda_graph_nops: int | None = None,
 ) -> TMeasurement:
    """
    Benchmark basic torch.mm as a roofline.
@@ -744,7 +745,7 @@ def use_cuda_graph_recommendation() -> str:
            """


-def print_timers(timers: list[TMeasurement], args: Optional[argparse.Namespace] = None):
+def print_timers(timers: list[TMeasurement], args: argparse.Namespace | None = None):
    compare = TBenchmark.Compare(timers)
    compare.print()

@@ -8,10 +8,9 @@ import math
 import os
 import pickle as pkl
 import time
-from collections.abc import Iterable
+from collections.abc import Callable, Iterable
 from dataclasses import dataclass
 from itertools import product
-from typing import Callable, Optional

 import pandas as pd
 import torch
@@ -63,23 +62,23 @@ class BenchmarkTensors:
    a: torch.Tensor

    w_q: torch.Tensor
-    group_size: Optional[int]
+    group_size: int | None
    wtype: ScalarType
    w_g_s: torch.Tensor
-    w_g_zp: Optional[torch.Tensor]
-    w_ch_s: Optional[torch.Tensor]
-    w_tok_s: Optional[torch.Tensor]
+    w_g_zp: torch.Tensor | None
+    w_ch_s: torch.Tensor | None
+    w_tok_s: torch.Tensor | None


@dataclass
 class TypeConfig:
    act_type: torch.dtype
    weight_type: ScalarType
-    output_type: Optional[torch.dtype]
-    group_scale_type: Optional[torch.dtype]
-    group_zero_type: Optional[torch.dtype]
-    channel_scale_type: Optional[torch.dtype]
-    token_scale_type: Optional[torch.dtype]
+    output_type: torch.dtype | None
+    group_scale_type: torch.dtype | None
+    group_zero_type: torch.dtype | None
+    channel_scale_type: torch.dtype | None
+    token_scale_type: torch.dtype | None


 def rand_data(shape, dtype=torch.float16, scale=1):
@@ -93,8 +92,8 @@ def quantize_and_pack(
    atype: torch.dtype,
    w: torch.Tensor,
    wtype: ScalarType,
-    stype: Optional[torch.dtype],
-    group_size: Optional[int],
+    stype: torch.dtype | None,
+    group_size: int | None,
    zero_points: bool = False,
 ):
    assert wtype.is_integer(), "TODO: support floating point weights"
@@ -113,7 +112,7 @@ def quantize_and_pack(


 def create_bench_tensors(
-    shape: tuple[int, int, int], types: TypeConfig, group_size: Optional[int]
+    shape: tuple[int, int, int], types: TypeConfig, group_size: int | None
 ) -> list[BenchmarkTensors]:
    m, n, k = shape

@@ -331,8 +330,8 @@ def bench_fns(label: str, sub_label: str, description: str, fns: list[Callable])
    return res


-_SWEEP_SCHEDULES_RESULTS: Optional[pd.DataFrame] = None
-_SWEEP_SCHEDULES_RESULTS_CSV: Optional[str] = None
+_SWEEP_SCHEDULES_RESULTS: pd.DataFrame | None = None
+_SWEEP_SCHEDULES_RESULTS_CSV: str | None = None


 def bench(
@@ -579,10 +579,12 @@ def main(args: argparse.Namespace):
        E = config.ffn_config.moe_num_experts
        topk = config.ffn_config.moe_top_k
        intermediate_size = config.ffn_config.ffn_hidden_size
+        hidden_size = config.hidden_size
    elif config.architectures[0] == "JambaForCausalLM":
        E = config.num_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.intermediate_size
+        hidden_size = config.hidden_size
    elif config.architectures[0] in (
        "DeepseekV2ForCausalLM",
        "DeepseekV3ForCausalLM",
@@ -592,6 +594,7 @@ def main(args: argparse.Namespace):
        E = config.n_routed_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.moe_intermediate_size
+        hidden_size = config.hidden_size
    elif config.architectures[0] in (
        "Qwen2MoeForCausalLM",
        "Qwen3MoeForCausalLM",
@@ -600,10 +603,18 @@ def main(args: argparse.Namespace):
        E = config.num_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.moe_intermediate_size
+        hidden_size = config.hidden_size
+    elif config.architectures[0] == "Qwen3VLMoeForConditionalGeneration":
+        text_config = config.get_text_config()
+        E = text_config.num_experts
+        topk = text_config.num_experts_per_tok
+        intermediate_size = text_config.moe_intermediate_size
+        hidden_size = text_config.hidden_size
    elif config.architectures[0] in ("HunYuanMoEV1ForCausalLM"):
        E = config.num_experts
        topk = config.moe_topk[0]
        intermediate_size = config.moe_intermediate_size[0]
+        hidden_size = config.hidden_size
    else:
        # Support for llama4
        config = config.get_text_config()
@@ -611,6 +622,7 @@ def main(args: argparse.Namespace):
        E = config.num_local_experts
        topk = config.num_experts_per_tok
        intermediate_size = config.intermediate_size
+        hidden_size = config.hidden_size
    enable_ep = bool(args.enable_expert_parallel)
    if enable_ep:
        ensure_divisibility(E, args.tp_size, "Number of experts")
@@ -619,7 +631,6 @@ def main(args: argparse.Namespace):
    else:
        ensure_divisibility(intermediate_size, args.tp_size, "intermediate_size")
        shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    hidden_size = config.hidden_size
    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
    use_fp8_w8a8 = args.dtype == "fp8_w8a8"
    use_int8_w8a16 = args.dtype == "int8_w8a16"
@@ -3,7 +3,6 @@

 import random
 import time
-from typing import Optional

 import torch

@@ -37,7 +36,7 @@ def main(
    seed: int,
    do_profile: bool,
    device: str = "cuda",
-    kv_cache_dtype: Optional[str] = None,
+    kv_cache_dtype: str | None = None,
 ) -> None:
    current_platform.seed_everything(seed)

@@ -3,8 +3,8 @@

 import argparse
 import math
+from collections.abc import Callable
 from contextlib import contextmanager
-from typing import Callable
 from unittest.mock import patch

 import torch
@@ -1,7 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from __future__ import annotations
-
 import random
 import time

@@ -1,7 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from __future__ import annotations
-
 import random
 import time

@@ -2,7 +2,6 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 import itertools
-from typing import Optional, Union

 import torch
 from flashinfer.norm import fused_add_rmsnorm, rmsnorm
@@ -21,8 +20,8 @@ class HuggingFaceRMSNorm(nn.Module):
    def forward(
        self,
        x: torch.Tensor,
-        residual: Optional[torch.Tensor] = None,
-    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
+        residual: torch.Tensor | None = None,
+    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        orig_dtype = x.dtype
        x = x.to(torch.float32)
        if residual is not None:
@@ -41,7 +40,7 @@ class HuggingFaceRMSNorm(nn.Module):
 def rmsnorm_naive(
    x: torch.Tensor,
    weight: torch.Tensor,
-    residual: Optional[torch.Tensor] = None,
+    residual: torch.Tensor | None = None,
    eps: float = 1e-6,
 ):
    naive_norm = HuggingFaceRMSNorm(x.shape[-1], eps=eps)
@@ -65,7 +64,7 @@ def rmsnorm_naive(
 def rmsnorm_flashinfer(
    x: torch.Tensor,
    weight: torch.Tensor,
-    residual: Optional[torch.Tensor] = None,
+    residual: torch.Tensor | None = None,
    eps: float = 1e-6,
 ):
    orig_shape = x.shape
@@ -89,7 +88,7 @@ def rmsnorm_flashinfer(
 def rmsnorm_vllm(
    x: torch.Tensor,
    weight: torch.Tensor,
-    residual: Optional[torch.Tensor] = None,
+    residual: torch.Tensor | None = None,
    eps: float = 1e-6,
 ):
    orig_shape = x.shape
@@ -2,7 +2,6 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 from itertools import accumulate
-from typing import Optional

 import nvtx
 import torch
@@ -18,7 +17,7 @@ def benchmark_rope_kernels_multi_lora(
    seq_len: int,
    num_heads: int,
    head_size: int,
-    rotary_dim: Optional[int],
+    rotary_dim: int | None,
    dtype: torch.dtype,
    seed: int,
    device: str,
@@ -1,5 +1,19 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+Comprehensive 3-way SiLU Benchmark Suite
+
+This benchmark compares three SiLU implementations:
+1. SiLU V2 (CUDA) - Optimized CUDA kernel implementation
+2. Triton Kernel - Triton-based implementation
+
+The suite generates detailed performance comparisons including:
+- Memory bandwidth utilization
+- Speedup ratios (baseline vs optimized implementations)
+- Performance across different expert configurations and token distributions
+"""
+
 from collections.abc import Callable

 import matplotlib.pyplot as plt
@@ -7,7 +21,7 @@ import numpy as np
 import torch

 from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
-    silu_mul_fp8_quant_deep_gemm_cuda,
+    persistent_masked_m_silu_mul_quant,
 )
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
@@ -94,6 +108,7 @@ def silu_mul_fp8_quant_deep_gemm_triton(
    num_parallel_tokens,
    group_size: int = 128,
    eps: float = 1e-10,
+    expert_offsets: torch.Tensor = None,
 ) -> tuple[torch.Tensor, torch.Tensor]:
    """Quantize silu(y[..., :H]) * y[..., H:] to FP8 with group per-token scales

@@ -174,7 +189,7 @@ def silu_mul_fp8_quant_deep_gemm_triton(


 # Parse generation strategies
-strategies = ["uniform", "max_t", "first_t"]
+strategies = ["random_imbalanced", "uniform", "max_t"]


 def benchmark(
@@ -195,15 +210,27 @@ def benchmark(
        current_platform.seed_everything(42 + seed_offset)
        y = torch.rand((E, T, 2 * H), dtype=torch.bfloat16, device="cuda").contiguous()

-        if gen_strategy == "uniform":
-            r = torch.rand(size=(E,), device="cuda")
+        if gen_strategy == "random_imbalanced":
+
+            def generate_expert_loads(n_e, total_tokens, ratio, device="cuda"):
+                mean = total_tokens // n_e
+                min_max = mean // ratio
+                e = torch.ones(size=(E,), dtype=torch.int64, device=device) * mean
+                e[0] = min_max
+                r = torch.rand(size=(E - 1,))
+                r /= r.sum()
+                r *= total_tokens - min_max
+                r = r.round().long()
+                e[1:] = r.to(device=device)
+                return e
+
+            tokens_per_expert = generate_expert_loads(E, total_tokens, 0.7, "cuda")
+        elif gen_strategy == "uniform":
+            r = torch.rand(size=(E,))
            r /= r.sum()
            r *= total_tokens
-            tokens_per_expert = r.int()
-            tokens_per_expert = torch.minimum(
-                tokens_per_expert,
-                torch.ones((E,), device=r.device, dtype=torch.int) * T,
-            )
+            r = r.round().long()
+            tokens_per_expert = r
        elif gen_strategy == "max_t":
            tokens_per_expert = torch.empty(size=(E,), dtype=torch.int32, device="cuda")
            tokens_per_expert.fill_(total_tokens / E)
@@ -281,40 +308,34 @@ def benchmark(


 def create_comparison_plot(
-    ratio, cuda_times, baseline_times, config_labels, strategy_name, id
+    ratios, silu_v2_times, triton_times, config_labels, strategy_name, id
 ):
-    """Create a comparison plot for a specific generation strategy"""
-    fig, ax = plt.subplots(1, 1, figsize=(16, 6))
+    fig, ax = plt.subplots(1, 1, figsize=(18, 6))

    # Configure x-axis positions
    x = np.arange(len(config_labels))
-    width = 0.35
+    width = 0.25

    # Execution Time plot (lower is better)
+    ax.bar(x, silu_v2_times, width, label="SiLU V2 (CUDA)", alpha=0.8, color="blue")
    ax.bar(
-        x - width / 2, cuda_times, width, label="CUDA Kernel", alpha=0.8, color="blue"
-    )
-    ax.bar(
-        x + width / 2,
-        baseline_times,
-        width,
-        label="Baseline",
-        alpha=0.8,
-        color="orange",
+        x + width, triton_times, width, label="Triton Kernel", alpha=0.8, color="green"
    )

-    # Add speedup labels over each bar pair
+    # Add speedup labels over each bar trio
    for i in range(len(x)):
-        speedup = ratio[i]
-        max_height = max(cuda_times[i], baseline_times[i])
+        triton_v2_speedup = ratios[i][1]  # triton/v2
+        max_height = max(silu_v2_times[i], triton_times[i])
+
+        # Triton/V2 speedup
        ax.text(
-            x[i],
+            x[i] + width / 2,
            max_height + max_height * 0.02,
-            f"{speedup:.2f}x",
+            f"{triton_v2_speedup:.2f}x",
            ha="center",
            va="bottom",
            fontweight="bold",
-            fontsize=9,
+            fontsize=8,
        )

    ax.set_xlabel("Configuration")
@@ -332,56 +353,75 @@ def create_comparison_plot(


 def create_combined_plot(all_results):
-    """Create a combined plot with all strategies in one PNG"""
    num_strategies = len(all_results)
-    fig, axes = plt.subplots(num_strategies, 1, figsize=(20, 6 * num_strategies))
+    fig, axes = plt.subplots(num_strategies, 1, figsize=(22, 7 * num_strategies))

    if num_strategies == 1:
        axes = [axes]

    for idx, (
        strategy_name,
-        ratio,
-        cuda_times,
-        baseline_times,
+        all_ratios,
+        all_silu_v2_results,
+        all_triton_results,
        config_labels,
+        config_x_axis,
    ) in enumerate(all_results):
        ax = axes[idx]

+        # Flatten the nested results to get bandwidth percentages for plotting
+        silu_v2_bandwidths = []
+        triton_bandwidths = []
+        flat_ratios = []
+
+        for config_results in all_silu_v2_results:
+            for result in config_results:
+                silu_v2_bandwidths.append(result[3])  # bandwidth percentage
+
+        for config_results in all_triton_results:
+            for result in config_results:
+                triton_bandwidths.append(result[3])  # bandwidth percentage
+
+        for config_ratios in all_ratios:
+            for ratio in config_ratios:
+                flat_ratios.append(ratio)
+
        # Configure x-axis positions
        x = np.arange(len(config_labels))
-        width = 0.35
+        width = 0.25

-        # Execution Time plot (lower is better)
+        # Bandwidth utilization plot (higher is better)
        ax.bar(
-            x - width / 2,
-            cuda_times,
+            x,
+            silu_v2_bandwidths,
            width,
-            label="CUDA Kernel",
+            label="SiLU V2 (CUDA)",
            alpha=0.8,
            color="blue",
        )
        ax.bar(
-            x + width / 2,
-            baseline_times,
+            x + width,
+            triton_bandwidths,
            width,
-            label="Baseline",
+            label="Triton Kernel",
            alpha=0.8,
-            color="orange",
+            color="green",
        )

-        # Add speedup labels over each bar pair
+        # Add speedup labels over each bar trio
        for i in range(len(x)):
-            speedup = ratio[i]
-            max_height = max(cuda_times[i], baseline_times[i])
+            triton_v2_speedup = flat_ratios[i]  # triton/v2
+            max_height = max(silu_v2_bandwidths[i], triton_bandwidths[i])
+
+            # Triton/V2 speedup
            ax.text(
-                x[i],
+                x[i] + width / 2,
                max_height + max_height * 0.02,
-                f"{speedup:.2f}x",
+                f"{triton_v2_speedup:.2f}x",
                ha="center",
                va="bottom",
                fontweight="bold",
-                fontsize=9,
+                fontsize=8,
            )

        ax.set_xlabel("Configuration")
@@ -395,7 +435,7 @@ def create_combined_plot(all_results):
        ax.grid(True, alpha=0.3)

    plt.tight_layout()
-    filename = "../../silu_bench/silu_benchmark_combined.png"
+    filename = "silu_benchmark_combined_3way.png"
    plt.savefig(filename, dpi=300, bbox_inches="tight")
    plt.show()

@@ -405,7 +445,9 @@ def create_combined_plot(all_results):
 outer_dim = 7168
 configs = [
    # DeepSeekV3 Configs
+    # (1, 56, 7168),
    (8, 1024, 7168),
+    # (32, 56, 7168),
    # DeepSeekV3 Configs
    (32, 1024, 7168),
    # DeepSeekV3 Configs
@@ -417,6 +459,7 @@ num_warmups = 20

 strategy_descriptions = {
    "uniform": "Uniform Random",
+    "random_imbalanced": "Imbalanced Random",
    "max_t": "Even Assignment",
    "first_t": "experts[0] = T, experts[1:] = 0",
 }
@@ -433,28 +476,31 @@ for id, strategy in enumerate(strategies):
    print(f"Testing strategy: {strategy_descriptions[strategy]}")
    print(f"{'=' * 60}")

-    # Collect benchmark data for both algorithms
+    # Collect benchmark data for all three algorithms
    config_labels = []
    config_x_axis = []
-    all_cuda_results = []
-    all_baseline_results = []
+    all_silu_v2_results = []
+    all_triton_results = []
    all_ratios = []

    for E, T, H in configs:
-        total_tokens_config = [8 * E, 16 * E, 32 * E, 64 * E, 128 * E, 256 * E]
+        total_tokens_config = []
+        for i in [8, 16, 32, 64, 128, 256, 512]:
+            if i <= T:
+                total_tokens_config.append(i * E)
        config_x_axis.append(total_tokens_config)

-        cuda_results = []
-        baseline_results = []
+        silu_v2_results = []
+        triton_results = []
        ratios = []

        for total_tokens in total_tokens_config:
            config_label = f"E={E},T={T},H={H},TT={total_tokens}"
            config_labels.append(config_label)

-            # CUDA kernel results
-            time_ms_cuda, gflops, gbps, perc = benchmark(
-                silu_mul_fp8_quant_deep_gemm_cuda,
+            # SiLU V2 (CUDA kernel) results
+            time_ms_silu_v2, gflops, gbps, perc = benchmark(
+                persistent_masked_m_silu_mul_quant,
                E,
                T,
                H,
@@ -463,9 +509,9 @@ for id, strategy in enumerate(strategies):
                num_warmups=num_warmups,
                gen_strategy=strategy,
            )
-            cuda_results.append((time_ms_cuda, gflops, gbps, perc))
+            silu_v2_results.append((time_ms_silu_v2, gflops, gbps, perc))

-            # Baseline results
+            # Triton kernel results
            time_ms_triton, gflops, gbps, perc = benchmark(
                silu_mul_fp8_quant_deep_gemm_triton,
                E,
@@ -476,12 +522,20 @@ for id, strategy in enumerate(strategies):
                num_warmups=num_warmups,
                gen_strategy=strategy,
            )
-            baseline_results.append((time_ms_triton, gflops, gbps, perc))
-            ratios.append(time_ms_triton / time_ms_cuda)
+            triton_results.append((time_ms_triton, gflops, gbps, perc))

-            print(f"Completed: {config_label}")
-        all_cuda_results.append(cuda_results)
-        all_baseline_results.append(baseline_results)
+            # Calculate speedup ratios (triton baseline / implementation)
+            triton_v2_ratio = time_ms_triton / time_ms_silu_v2
+            ratios.append(triton_v2_ratio)
+
+            print(
+                f"Completed: {config_label}:"
+                f" V2: {time_ms_silu_v2:.3f}ms,"
+                f" Triton: {time_ms_triton:.3f}ms"
+            )
+
+        all_silu_v2_results.append(silu_v2_results)
+        all_triton_results.append(triton_results)
        all_ratios.append(ratios)

    # Store results for combined plotting
@@ -489,8 +543,8 @@ for id, strategy in enumerate(strategies):
        (
            strategy_descriptions[strategy],
            all_ratios,
-            all_cuda_results,
-            all_baseline_results,
+            all_silu_v2_results,
+            all_triton_results,
            config_labels,
            config_x_axis,
        )
@@ -498,15 +552,18 @@ for id, strategy in enumerate(strategies):

    # Print summary table for this strategy
    print(f"\nSummary Table - {strategy_descriptions[strategy]}:")
-    print(f"{'Config':<20} {'CUDA Time(ms)':<12} {'Base Time(ms)':<12} {'Speedup':<8}")
-    print("-" * 60)
+    print(f" {'V2 Time(ms)':<12} {'Triton Time(ms)':<14} {'Triton/V2':<10}")
+    print("-" * 90)

    for i, (E, T, H) in enumerate(configs):
-        speedup = baseline_results[i][0] / cuda_results[i][0]
+        # Get the first result for each config (simplifying for summary)
+        v2_time = silu_v2_results[i][0]
+        triton_time = triton_results[i][0]
+        triton_v2_speedup = triton_time / v2_time
        config_label = f"E={E:3d},T={T:4d},H={H:4d}"
        print(
-            f"{config_label:<20} {cuda_results[i][0]:8.5f} "
-            f"{baseline_results[i][0]:8.5f} {speedup:6.2f}x"
+            f"{config_label:<20} {v2_time:8.5f} {triton_time:10.5f} "
+            f"{triton_v2_speedup:8.2f}x"
        )


@@ -514,15 +571,14 @@ def create_total_tokens_plot(all_results):
    num_strategies = len(all_results)
    num_configs = len(configs)

-    # Create side-by-side subplots: 2 columns for speedup and bandwidth percentage
    fig, axs = plt.subplots(
-        num_strategies, num_configs * 2, figsize=(28, 6 * num_strategies)
+        num_strategies, num_configs * 2, figsize=(32, 8 * num_strategies)
    )

    # Add main title to the entire figure
    fig.suptitle(
-        "Performance Analysis: Speedup vs Bandwidth Utilization (Triton & CUDA)",
-        fontsize=16,
+        "Performance Analysis: Speedup vs Bandwidth Utilization (SiLU V2, and Triton)",
+        fontsize=18,
        fontweight="bold",
        y=0.98,
    )
@@ -539,8 +595,8 @@ def create_total_tokens_plot(all_results):
        (
            strategy_name,
            all_ratios,
-            all_cuda_results,
-            all_baseline_results,
+            all_silu_v2_results,
+            all_triton_results,
            config_labels,
            config_x_axis,
        ) = result
@@ -555,42 +611,54 @@ def create_total_tokens_plot(all_results):
            ratios = all_ratios[config_idx]
            total_tokens_values = config_x_axis[config_idx]

-            # Extract CUDA and Triton bandwidth percentages
-            cuda_bandwidth_percentages = [
-                result[3] for result in all_cuda_results[config_idx]
+            # Extract speedup ratios
+            triton_v2_ratios = [ratio for ratio in ratios]
+
+            # Extract bandwidth percentages for all implementations
+            v2_bandwidth_percentages = [
+                result[3] for result in all_silu_v2_results[config_idx]
            ]
            triton_bandwidth_percentages = [
-                result[3] for result in all_baseline_results[config_idx]
+                result[3] for result in all_triton_results[config_idx]
            ]

            # Plot speedup ratios vs total tokens (left plot)
            ax_speedup.plot(
-                total_tokens_values, ratios, "bo-", linewidth=3, markersize=8
+                total_tokens_values,
+                triton_v2_ratios,
+                "go-",
+                linewidth=3,
+                markersize=8,
+                label="Triton/V2 Speedup",
            )
            ax_speedup.set_title(
-                f"{strategy_name}\nSpeedup (CUDA/Triton)\nE={E}, T={T}, H={H}",
+                f"{strategy_name}\nSpeedup vs Baseline (Triton)\nE={E}, T={T}, H={H}",
                fontsize=12,
                fontweight="bold",
            )
            ax_speedup.set_xlabel("Total Tokens", fontweight="bold", fontsize=11)
            ax_speedup.set_ylabel("Speedup Ratio", fontweight="bold", fontsize=11)
+            ax_speedup.legend(prop={"weight": "bold"})
            ax_speedup.grid(True, alpha=0.3)

+            # Plot bandwidth utilization (right plot)
            ax_bandwidth.plot(
                total_tokens_values,
-                cuda_bandwidth_percentages,
-                "ro-",
+                v2_bandwidth_percentages,
+                "o-",
                linewidth=3,
                markersize=8,
-                label="CUDA",
+                label="SiLU V2",
+                color="blue",
            )
            ax_bandwidth.plot(
                total_tokens_values,
                triton_bandwidth_percentages,
-                "go-",
+                "o-",
                linewidth=3,
                markersize=8,
                label="Triton",
+                color="green",
            )
            ax_bandwidth.set_title(
                f"{strategy_name}\nBandwidth Utilization (Hopper)\nE={E}, T={T}, H={H}",
@@ -618,38 +686,12 @@ def create_total_tokens_plot(all_results):
                for label in ax.get_xticklabels() + ax.get_yticklabels():
                    label.set_fontweight("bold")

-            # Add value labels on speedup points
-            for x, y in zip(total_tokens_values, ratios):
+            # Add value labels on Triton/V2 speedup points
+            for x, y in zip(total_tokens_values, triton_v2_ratios):
                ax_speedup.annotate(
                    f"{y:.2f}x",
                    (x, y),
                    textcoords="offset points",
-                    xytext=(0, 12),
-                    ha="center",
-                    fontsize=10,
-                    fontweight="bold",
-                    bbox=dict(boxstyle="round,pad=0.3", facecolor="white", alpha=0.7),
-                )
-
-            # Add value labels on CUDA bandwidth points
-            for x, y in zip(total_tokens_values, cuda_bandwidth_percentages):
-                ax_bandwidth.annotate(
-                    f"{y:.1f}%",
-                    (x, y),
-                    textcoords="offset points",
-                    xytext=(0, 12),
-                    ha="center",
-                    fontsize=9,
-                    fontweight="bold",
-                    bbox=dict(boxstyle="round,pad=0.2", facecolor="red", alpha=0.3),
-                )
-
-            # Add value labels on Triton bandwidth points
-            for x, y in zip(total_tokens_values, triton_bandwidth_percentages):
-                ax_bandwidth.annotate(
-                    f"{y:.1f}%",
-                    (x, y),
-                    textcoords="offset points",
                    xytext=(0, -15),
                    ha="center",
                    fontsize=9,
@@ -659,17 +701,20 @@ def create_total_tokens_plot(all_results):

    plt.tight_layout()
    plt.subplots_adjust(top=0.93)  # Make room for main title
-    filename = "silu_benchmark_total_tokens.png"
+    filename = "silu_benchmark_total_tokens_3way.png"
    plt.savefig(filename, dpi=300, bbox_inches="tight")
    plt.show()

    return filename


-# Create combined plot with all strategies
-combined_plot_filename = create_total_tokens_plot(all_results)
+# Create comprehensive 3-way comparison plots
+combined_plot_filename = create_combined_plot(all_results)
+total_tokens_plot_filename = create_total_tokens_plot(all_results)

-print(f"\n{'=' * 60}")
-print("Benchmark Complete!")
-print(f"Generated combined plot: {combined_plot_filename}")
-print(f"{'=' * 60}")
+print(f"\n{'=' * 80}")
+print("3-Way Benchmark Suite Complete!")
+print(f"Generated combined comparison plot: {combined_plot_filename}")
+print(f"Generated total tokens analysis plot: {total_tokens_plot_filename}")
+print("Compared: SiLU V2 (CUDA), and Triton implementations")
+print(f"{'=' * 80}")
@@ -4,7 +4,6 @@
 import csv
 import os
 from datetime import datetime
-from typing import Optional

 import flashinfer
 import torch
@@ -28,9 +27,7 @@ def to_float8(x, dtype=torch.float8_e4m3fn):
@torch.no_grad()
 def benchmark_decode(
    dtype: torch.dtype,
-    quant_dtypes: tuple[
-        Optional[torch.dtype], Optional[torch.dtype], Optional[torch.dtype]
-    ],
+    quant_dtypes: tuple[torch.dtype | None, torch.dtype | None, torch.dtype | None],
    batch_size: int,
    max_seq_len: int,
    num_heads: tuple[int, int] = (64, 8),
@@ -4,7 +4,6 @@
 import csv
 import os
 from datetime import datetime
-from typing import Optional

 import flashinfer
 import torch
@@ -28,9 +27,7 @@ def to_float8(x, dtype=torch.float8_e4m3fn):
@torch.no_grad()
 def benchmark_prefill(
    dtype: torch.dtype,
-    quant_dtypes: tuple[
-        Optional[torch.dtype], Optional[torch.dtype], Optional[torch.dtype]
-    ],
+    quant_dtypes: tuple[torch.dtype | None, torch.dtype | None, torch.dtype | None],
    batch_size: int,
    max_seq_len: int,
    num_heads: tuple[int, int] = (64, 8),
@@ -14,7 +14,7 @@ import torch
 from tqdm import tqdm

 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    _w8a8_block_fp8_matmul,
+    _w8a8_triton_block_scaled_mm,
 )
 from vllm.platforms import current_platform
 from vllm.triton_utils import triton
@@ -83,7 +83,7 @@ def w8a8_block_matmul(
        )

    if A.dtype == torch.float8_e4m3fn:
-        kernel = _w8a8_block_fp8_matmul
+        kernel = _w8a8_triton_block_scaled_mm
    else:
        raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.")

@@ -2,8 +2,8 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 import dataclasses
-from collections.abc import Iterable
-from typing import Any, Callable, Optional
+from collections.abc import Callable, Iterable
+from typing import Any

 import torch
 import torch.utils.benchmark as TBenchmark
@@ -55,7 +55,7 @@ class Bench:

    def __init__(
        self,
-        cuda_graph_params: Optional[CudaGraphBenchParams],
+        cuda_graph_params: CudaGraphBenchParams | None,
        label: str,
        sub_label: str,
        description: str,
@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 from abc import ABC, abstractmethod
 from statistics import mean
-from typing import Any, NamedTuple, Optional, Union
+from typing import Any, NamedTuple

 import numpy as np  # type: ignore
 import pandas as pd  # type: ignore
@@ -35,8 +35,8 @@ class Distribution(ABC):
 class UniformDistribution(Distribution):
    def __init__(
        self,
-        min_val: Union[int, float],
-        max_val: Union[int, float],
+        min_val: int | float,
+        max_val: int | float,
        is_integer: bool = True,
    ) -> None:
        self.min_val = min_val
@@ -56,7 +56,7 @@ class UniformDistribution(Distribution):


 class ConstantDistribution(Distribution):
-    def __init__(self, value: Union[int, float]) -> None:
+    def __init__(self, value: int | float) -> None:
        self.value = value
        self.max_val = value

@@ -68,7 +68,7 @@ class ConstantDistribution(Distribution):


 class ZipfDistribution(Distribution):
-    def __init__(self, alpha: float, max_val: Optional[int] = None) -> None:
+    def __init__(self, alpha: float, max_val: int | None = None) -> None:
        self.alpha = alpha
        self.max_val = max_val

@@ -83,7 +83,7 @@ class ZipfDistribution(Distribution):


 class PoissonDistribution(Distribution):
-    def __init__(self, alpha: float, max_val: Optional[int] = None) -> None:
+    def __init__(self, alpha: float, max_val: int | None = None) -> None:
        self.alpha = alpha
        self.max_val = max_val

@@ -100,11 +100,11 @@ class PoissonDistribution(Distribution):
 class LognormalDistribution(Distribution):
    def __init__(
        self,
-        mean: Optional[float] = None,
-        sigma: Optional[float] = None,
-        average: Optional[int] = None,
-        median_ratio: Optional[float] = None,
-        max_val: Optional[int] = None,
+        mean: float | None = None,
+        sigma: float | None = None,
+        average: int | None = None,
+        median_ratio: float | None = None,
+        max_val: int | None = None,
    ) -> None:
        self.average = average
        self.median_ratio = median_ratio
@@ -13,7 +13,7 @@ from datetime import datetime
 from enum import Enum
 from http import HTTPStatus
 from statistics import mean
-from typing import NamedTuple, Optional, Union
+from typing import NamedTuple

 import aiohttp  # type: ignore
 import numpy as np  # type: ignore
@@ -46,9 +46,9 @@ class ConversationSampling(str, Enum):

 class ClientArgs(NamedTuple):
    seed: int
-    max_num_requests: Optional[int]
+    max_num_requests: int | None
    skip_first_turn: bool
-    max_turns: Optional[int]
+    max_turns: int | None
    max_active_conversations: int
    verbose: bool
    print_content: bool
@@ -109,9 +109,9 @@ class RequestStats(NamedTuple):

 class MetricStats:
    def __init__(self) -> None:
-        self.min: Optional[float] = None
-        self.max: Optional[float] = None
-        self.avg: Optional[float] = None
+        self.min: float | None = None
+        self.max: float | None = None
+        self.avg: float | None = None
        self.sum = 0.0
        self.count = 0

@@ -143,7 +143,7 @@ class MovingAverage:
        self.index = 0
        self.sum = 0.0
        self.count = 0
-        self.avg: Optional[float] = None
+        self.avg: float | None = None

    def update(self, new_value: float) -> None:
        if self.count < self.window_size:
@@ -169,7 +169,7 @@ class MovingAverage:
 class DebugStats:
    def __init__(self, logger: logging.Logger, window_size: int) -> None:
        self.logger = logger
-        self.metrics: dict[str, Union[MovingAverage, MetricStats]] = {
+        self.metrics: dict[str, MovingAverage | MetricStats] = {
            "moving_avg_ttft_ms": MovingAverage(window_size),
            "moving_avg_tpot_ms": MovingAverage(window_size),
            "ttft_ms": MetricStats(),
@@ -198,14 +198,6 @@ class DebugStats:
        self.logger.info("-" * 50)


-# Must support Python 3.8, we can't use str.removeprefix(prefix)
-# introduced in Python 3.9
-def remove_prefix(text: str, prefix: str) -> str:
-    if text.startswith(prefix):
-        return text[len(prefix) :]
-    return text
-
-
 def nanosec_to_millisec(value: float) -> float:
    return value / 1000000.0

@@ -220,8 +212,8 @@ async def send_request(
    chat_url: str,
    model: str,
    stream: bool = True,
-    min_tokens: Optional[int] = None,
-    max_tokens: Optional[int] = None,
+    min_tokens: int | None = None,
+    max_tokens: int | None = None,
 ) -> ServerResponse:
    payload = {
        "model": model,
@@ -250,9 +242,9 @@ async def send_request(
    timeout = aiohttp.ClientTimeout(total=timeout_sec)

    valid_response = True
-    ttft: Optional[float] = None
+    ttft: float | None = None
    chunk_delay: list[int] = []
-    latency: Optional[float] = None
+    latency: float | None = None
    first_chunk = ""
    generated_text = ""

@@ -269,7 +261,7 @@ async def send_request(
                if not chunk_bytes:
                    continue

-                chunk = remove_prefix(chunk_bytes.decode("utf-8"), "data: ")
+                chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
                if chunk == "[DONE]":
                    # End of stream
                    latency = time.perf_counter_ns() - start_time
@@ -364,7 +356,7 @@ async def send_turn(
    req_args: RequestArgs,
    verbose: bool,
    verify_output: bool,
-) -> Optional[RequestStats]:
+) -> RequestStats | None:
    assert messages_to_use > 0
    assert messages_to_use <= len(conversation_messages)

@@ -644,7 +636,7 @@ async def client_main(

            if args.verbose:
                curr_time_sec: float = time.perf_counter()
-                time_since_last_turn: Union[str, float] = "N/A"
+                time_since_last_turn: str | float = "N/A"
                if conv_id in time_of_last_turn:
                    time_since_last_turn = round(
                        curr_time_sec - time_of_last_turn[conv_id], 3
@@ -769,7 +761,7 @@ def get_client_config(
            "Number of conversations must be equal or larger than the number of clients"
        )

-    max_req_per_client: Optional[int] = None
+    max_req_per_client: int | None = None
    if args.max_num_requests is not None:
        # Max number of requests per client
        req_per_client = args.max_num_requests // args.num_clients
@@ -936,13 +928,13 @@ async def main_mp(
                    f"{num_clients_finished} out of {bench_args.num_clients} clients finished, collected {len(client_metrics)} measurements, runtime {runtime_sec:.3f} sec{Color.RESET}"  # noqa: E501
                )

-                rps: Union[str, float] = round(len(client_metrics) / runtime_sec, 3)
+                rps: str | float = round(len(client_metrics) / runtime_sec, 3)
                if len(client_metrics) < (5 * bench_args.num_clients):
                    # Do not estimate the RPS if the number of samples is very low
                    # (threshold can be tuned if needed)
                    rps = "N/A"

-                runtime_left_sec: Union[str, float] = round(
+                runtime_left_sec: str | float = round(
                    (runtime_sec / finished_convs) * (total_convs - finished_convs), 3
                )
                if percent < 0.05:
@@ -1032,7 +1024,7 @@ def process_statistics(
    warmup_percentages: list[float],
    test_params: dict,
    verbose: bool,
-    gen_conv_args: Optional[GenConvArgs] = None,
+    gen_conv_args: GenConvArgs | None = None,
    excel_output: bool = False,
 ) -> None:
    if len(client_metrics) == 0:
@@ -13,7 +13,7 @@ import argparse
 import json
 import random
 from statistics import mean
-from typing import Any, Optional
+from typing import Any

 import pandas as pd  # type: ignore
 import tqdm  # type: ignore
@@ -25,7 +25,7 @@ def has_non_english_chars(text: str) -> bool:


 def content_is_valid(
-    content: str, min_content_len: Optional[int], max_content_len: Optional[int]
+    content: str, min_content_len: int | None, max_content_len: int | None
 ) -> bool:
    if min_content_len and len(content) < min_content_len:
        return False
@@ -37,7 +37,7 @@ def content_is_valid(


 def print_stats(
-    conversations: "list[dict[Any, Any]]", tokenizer: Optional[AutoTokenizer] = None
+    conversations: "list[dict[Any, Any]]", tokenizer: AutoTokenizer | None = None
 ) -> None:
    # Collect statistics
    stats = []
@@ -109,12 +109,12 @@ def convert_sharegpt_to_openai(
    seed: int,
    input_file: str,
    output_file: str,
-    max_items: Optional[int],
-    min_content_len: Optional[int] = None,
-    max_content_len: Optional[int] = None,
-    min_turns: Optional[int] = None,
-    max_turns: Optional[int] = None,
-    model: Optional[str] = None,
+    max_items: int | None,
+    min_content_len: int | None = None,
+    max_content_len: int | None = None,
+    min_turns: int | None = None,
+    max_turns: int | None = None,
+    model: str | None = None,
 ) -> None:
    if min_turns and max_turns:
        assert min_turns <= max_turns
@@ -198,13 +198,24 @@ else()
 endif()

 if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
-    FetchContent_Declare(
-        oneDNN
-        GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-        GIT_TAG v3.9
-        GIT_PROGRESS TRUE
-        GIT_SHALLOW TRUE
-    )
+    set(FETCHCONTENT_SOURCE_DIR_ONEDNN "$ENV{FETCHCONTENT_SOURCE_DIR_ONEDNN}" CACHE PATH "Path to a local oneDNN source directory.")
+
+    if(FETCHCONTENT_SOURCE_DIR_ONEDNN)
+        message(STATUS "Using oneDNN from specified source directory: ${FETCHCONTENT_SOURCE_DIR_ONEDNN}")
+        FetchContent_Declare(
+            oneDNN
+            SOURCE_DIR ${FETCHCONTENT_SOURCE_DIR_ONEDNN}
+        )
+    else()
+        message(STATUS "Downloading oneDNN from GitHub")
+        FetchContent_Declare(
+            oneDNN
+            GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
+            GIT_TAG v3.9
+            GIT_PROGRESS TRUE
+            GIT_SHALLOW TRUE
+        )
+    endif()

    if(USE_ACL)
        find_library(ARM_COMPUTE_LIBRARY NAMES arm_compute PATHS $ENV{ACL_ROOT_DIR}/build/)
@@ -227,7 +238,7 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
    set(ONEDNN_ENABLE_ITT_TASKS "OFF")
    set(ONEDNN_ENABLE_MAX_CPU_ISA "OFF")
    set(ONEDNN_ENABLE_CPU_ISA_HINTS "OFF")
-    set(ONEDNN_VERBOSE "ON")
+    set(ONEDNN_VERBOSE "OFF")
    set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)

    FetchContent_MakeAvailable(oneDNN)
@@ -309,4 +320,4 @@ define_gpu_extension_target(
    WITH_SOABI
 )

-message(STATUS "Enabling C extension.")
+message(STATUS "Enabling C extension.")
@@ -0,0 +1,97 @@
+include(FetchContent)
+
+set(CUTLASS_INCLUDE_DIR "${CUTLASS_INCLUDE_DIR}" CACHE PATH "Path to CUTLASS include/ directory")
+
+if(DEFINED ENV{QUTLASS_SRC_DIR})
+  set(QUTLASS_SRC_DIR $ENV{QUTLASS_SRC_DIR})
+endif()
+
+if(QUTLASS_SRC_DIR)
+  FetchContent_Declare(
+    qutlass
+    SOURCE_DIR ${QUTLASS_SRC_DIR}
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND ""
+  )
+else()
+  FetchContent_Declare(
+    qutlass
+    GIT_REPOSITORY https://github.com/IST-DASLab/qutlass.git
+    GIT_TAG 830d2c4537c7396e14a02a46fbddd18b5d107c65
+    GIT_PROGRESS TRUE
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND ""
+  )
+  FetchContent_Populate(qutlass)
+  set(qutlass_SOURCE_DIR "${qutlass_SOURCE_DIR}")
+endif()
+
+if(NOT qutlass_SOURCE_DIR)
+  message(FATAL_ERROR "[QUTLASS] source directory could not be resolved.")
+endif()
+message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}")
+
+cuda_archs_loose_intersection(QUTLASS_ARCHS "12.0a;10.0a" "${CUDA_ARCHS}")
+if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND QUTLASS_ARCHS)
+
+  if(QUTLASS_ARCHS MATCHES "10\\.0a")
+    set(QUTLASS_TARGET_CC 100)
+  elseif(QUTLASS_ARCHS MATCHES "12\\.0a")
+    set(QUTLASS_TARGET_CC 120)
+  else()
+    message(FATAL_ERROR "[QUTLASS] internal error parsing CUDA_ARCHS='${QUTLASS_ARCHS}'.")
+  endif()
+
+  set(QUTLASS_SOURCES
+    ${qutlass_SOURCE_DIR}/qutlass/csrc/bindings.cpp
+    ${qutlass_SOURCE_DIR}/qutlass/csrc/gemm.cu
+    ${qutlass_SOURCE_DIR}/qutlass/csrc/gemm_ada.cu
+    ${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_mx.cu
+    ${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_nv.cu
+    ${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_mx_sm100.cu
+    ${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_nv_sm100.cu
+  )
+
+  set(QUTLASS_INCLUDES
+    ${qutlass_SOURCE_DIR}
+    ${qutlass_SOURCE_DIR}/qutlass
+    ${qutlass_SOURCE_DIR}/qutlass/csrc/include
+    ${qutlass_SOURCE_DIR}/qutlass/csrc/include/cutlass_extensions
+  )
+
+  if(CUTLASS_INCLUDE_DIR AND EXISTS "${CUTLASS_INCLUDE_DIR}/cutlass/cutlass.h")
+    list(APPEND QUTLASS_INCLUDES "${CUTLASS_INCLUDE_DIR}")
+  elseif(EXISTS "${qutlass_SOURCE_DIR}/qutlass/third_party/cutlass/include/cutlass/cutlass.h")
+    list(APPEND QUTLASS_INCLUDES "${qutlass_SOURCE_DIR}/qutlass/third_party/cutlass/include")
+    message(STATUS "[QUTLASS] Using QuTLASS vendored CUTLASS headers (no vLLM CUTLASS detected).")
+  else()
+    message(FATAL_ERROR "[QUTLASS] CUTLASS headers not found. "
+                        "Set -DCUTLASS_INCLUDE_DIR=/path/to/cutlass/include")
+  endif()
+
+  set_gencode_flags_for_srcs(
+    SRCS "${QUTLASS_SOURCES}"
+    CUDA_ARCHS "${QUTLASS_ARCHS}"
+  )
+
+  target_sources(_C PRIVATE ${QUTLASS_SOURCES})
+  target_include_directories(_C PRIVATE ${QUTLASS_INCLUDES})
+  target_compile_definitions(_C PRIVATE
+    QUTLASS_DISABLE_PYBIND=1
+    TARGET_CUDA_ARCH=${QUTLASS_TARGET_CC}
+  )
+
+  set_property(SOURCE ${QUTLASS_SOURCES} APPEND PROPERTY COMPILE_OPTIONS
+    $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr --use_fast_math -O3>
+  )
+
+else()
+  if("${CMAKE_CUDA_COMPILER_VERSION}" VERSION_LESS "12.8")
+    message(STATUS
+      "[QUTLASS] Skipping build: CUDA 12.8 or newer is required (found ${CMAKE_CUDA_COMPILER_VERSION}).")
+  else()
+    message(STATUS
+      "[QUTLASS] Skipping build: no supported arch (12.0a / 10.0a) found in "
+      "CUDA_ARCHS='${CUDA_ARCHS}'.")
+  endif()
+endif()
@@ -38,7 +38,7 @@ else()
  FetchContent_Declare(
          vllm-flash-attn
          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 4695e6bed5366c41e28c06cd86170166e4f43d00
+          GIT_TAG 8f468e7da54a8e2f98abfa7c38636aac91c0cba1
          GIT_PROGRESS TRUE
          # Don't share the vllm-flash-attn build between build types
          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
@@ -28,10 +28,10 @@

 #ifdef USE_ROCM
  #include <hip/hip_bf16.h>
-  #include "../quantization/fp8/amd/quant_utils.cuh"
+  #include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
 typedef __hip_bfloat16 __nv_bfloat16;
 #else
-  #include "../quantization/fp8/nvidia/quant_utils.cuh"
+  #include "../quantization/w8a8/fp8/nvidia/quant_utils.cuh"
 #endif

 #define MAX(a, b) ((a) > (b) ? (a) : (b))
@@ -64,3 +64,11 @@ void indexer_k_quant_and_cache(
    torch::Tensor& slot_mapping,  // [num_tokens]
    int64_t quant_block_size,     // quantization block size
    const std::string& scale_fmt);
+
+// Extract function to gather quantized K cache
+void cp_gather_indexer_k_quant_cache(
+    const torch::Tensor& kv_cache,  // [num_blocks, block_size, cache_stride]
+    torch::Tensor& dst_k,           // [num_tokens, head_dim]
+    torch::Tensor& dst_scale,  // [num_tokens, head_dim / quant_block_size * 4]
+    const torch::Tensor& block_table,   // [batch_size, num_blocks]
+    const torch::Tensor& cu_seq_lens);  // [batch_size + 1]
@@ -9,9 +9,9 @@
 #include "quantization/vectorization_utils.cuh"

 #ifdef USE_ROCM
-  #include "quantization/fp8/amd/quant_utils.cuh"
+  #include "quantization/w8a8/fp8/amd/quant_utils.cuh"
 #else
-  #include "quantization/fp8/nvidia/quant_utils.cuh"
+  #include "quantization/w8a8/fp8/nvidia/quant_utils.cuh"
 #endif

 #include <algorithm>
@@ -572,6 +572,70 @@ __global__ void indexer_k_quant_and_cache_kernel(
  }
 }

+template <int BLOCK_Y_SIZE>
+__global__ void cp_gather_indexer_k_quant_cache_kernel(
+    const char* __restrict__ kv_cache,  // [num_blocks, block_size,
+                                        // cache_stride]
+    char* __restrict__ dst_k,           // [num_tokens, head_dim]
+    char* __restrict__ dst_scale,  // [num_tokens, head_dim / quant_block_size *
+                                   // 4]
+    const int* __restrict__ block_table,  // [batch_size, num_blocks]
+    const int* __restrict__ cu_seq_lens,  // [batch_size + 1]
+    const int batch_size,                 // batch size
+    const int64_t token_stride,           // stride for each token in dst_k
+    const int64_t head_dim,               // dimension of each head
+    const int64_t block_stride,           // stride for each block in kv_cache
+    const int64_t cache_token_stride,     // stride for each token in kv_cache
+    const int64_t cache_block_size,  // num_tokens for each block in kv_cache
+    const int num_blocks,            // number of blocks
+    const int num_tokens,            // number of tokens
+    const int quant_block_size       // quantization block size
+) {
+  constexpr int VEC_SIZE = sizeof(float4) / sizeof(char);
+  const int token_idx = blockIdx.x * blockDim.y + threadIdx.y;
+  const int head_idx = (blockIdx.y * blockDim.x + threadIdx.x) * VEC_SIZE;
+  // Find batch index within a block
+  __shared__ int batch_idx[BLOCK_Y_SIZE];
+  for (int iter = 0; iter < cuda_utils::ceil_div(batch_size, int(blockDim.x));
+       iter++) {
+    int tid = iter * blockDim.x + threadIdx.x;
+    if (tid < batch_size) {
+      const int seq_start = cu_seq_lens[tid];
+      const int seq_end = cu_seq_lens[tid + 1];
+      if (token_idx >= seq_start && token_idx < seq_end) {
+        batch_idx[threadIdx.y] = tid;
+      }
+    }
+  }
+
+#ifndef USE_ROCM
+  __syncwarp();
+#endif
+
+  if (head_idx >= head_dim || token_idx >= num_tokens) {
+    return;
+  }
+  const int inbatch_seq_idx = token_idx - cu_seq_lens[batch_idx[threadIdx.y]];
+  const int block_idx = block_table[batch_idx[threadIdx.y] * num_blocks +
+                                    inbatch_seq_idx / cache_block_size];
+  const int64_t src_block_offset = block_idx * block_stride;
+  const int64_t cache_inblock_offset =
+      (inbatch_seq_idx % cache_block_size) * head_dim + head_idx;
+  const int64_t src_inblock_offset = src_block_offset + cache_inblock_offset;
+  const int64_t dst_inblock_offset = token_idx * token_stride + head_idx;
+
+  reinterpret_cast<float4*>(dst_k)[dst_inblock_offset / VEC_SIZE] =
+      reinterpret_cast<const float4*>(kv_cache)[src_inblock_offset / VEC_SIZE];
+  ;
+  if (threadIdx.x == 0) {
+    const int64_t src_scale_offset =
+        src_block_offset + cache_block_size * head_dim +
+        cache_inblock_offset * 4 / quant_block_size;
+    reinterpret_cast<float*>(dst_scale)[dst_inblock_offset / quant_block_size] =
+        reinterpret_cast<const float*>(kv_cache)[src_scale_offset / 4];
+  }
+}
+
 }  // namespace vllm

 // KV_T is the data type of key and value tensors.
@@ -1173,3 +1237,59 @@ void indexer_k_quant_and_cache(
  DISPATCH_BY_KV_CACHE_DTYPE(k.dtype(), "fp8_e4m3",
                             CALL_INDEXER_K_QUANT_AND_CACHE);
 }
+
+// Macro to dispatch the kernel based on the data amount.
+#define CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(BLOCK_Y_SIZE)                  \
+  vllm::cp_gather_indexer_k_quant_cache_kernel<BLOCK_Y_SIZE>                \
+      <<<dim3((num_tokens + BLOCK_Y_SIZE - 1) / BLOCK_Y_SIZE,               \
+              (head_dim + 8 * vec_size - 1) / (8 * vec_size)),              \
+         dim3(8, BLOCK_Y_SIZE), 0, stream>>>(                               \
+          reinterpret_cast<char*>(kv_cache.data_ptr()),                     \
+          reinterpret_cast<char*>(dst_k.data_ptr()),                        \
+          reinterpret_cast<char*>(dst_scale.data_ptr()),                    \
+          block_table.data_ptr<int32_t>(), cu_seq_lens.data_ptr<int32_t>(), \
+          batch_size, dst_k.stride(0), dst_k.size(1), kv_cache.stride(0),   \
+          kv_cache.stride(1), kv_cache.size(1), block_table.size(1),        \
+          num_tokens, quant_block_size);
+
+void cp_gather_indexer_k_quant_cache(
+    const torch::Tensor& kv_cache,  // [num_blocks, block_size, cache_stride]
+    torch::Tensor& dst_k,           // [num_tokens, head_dim]
+    torch::Tensor& dst_scale,  // [num_tokens, head_dim / quant_block_size * 4]
+    const torch::Tensor& block_table,  // [batch_size, num_blocks]
+    const torch::Tensor& cu_seq_lens   // [batch_size + 1]
+) {
+  int batch_size = block_table.size(0);
+  int num_tokens = dst_k.size(0);
+  int head_dim = dst_k.size(1);
+  int quant_block_size = head_dim * 4 / dst_scale.size(1);
+
+  TORCH_CHECK(kv_cache.device() == dst_k.device(),
+              "kv_cache and dst_k must be on the same device");
+  TORCH_CHECK(kv_cache.device() == dst_scale.device(),
+              "kv_cache and dst_scale must be on the same device");
+  TORCH_CHECK(kv_cache.device() == block_table.device(),
+              "kv_cache and block_table must be on the same device");
+  TORCH_CHECK(kv_cache.device() == cu_seq_lens.device(),
+              "kv_cache and cu_seq_lens must be on the same device");
+  TORCH_CHECK(head_dim % quant_block_size == 0,
+              "head_dim must be divisible by quant_block_size");
+
+  constexpr int vec_size = 16;
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(kv_cache));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  if (num_tokens < 32) {
+    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(1);
+  } else if (num_tokens < 64) {
+    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(2);
+  } else if (num_tokens < 128) {
+    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(4);
+  } else if (num_tokens < 256) {
+    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(8);
+  } else if (num_tokens < 512) {
+    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(16);
+  } else {
+    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(32);
+  }
+}
@@ -8,9 +8,12 @@ namespace vllm {
 // vllm_kernel_override_batch_invariant(); returns true
 // if env VLLM_KERNEL_OVERRIDE_BATCH_INVARIANT=1
 inline bool vllm_kernel_override_batch_invariant() {
-  std::string env_key = "VLLM_KERNEL_OVERRIDE_BATCH_INVARIANT";
-  const char* val = std::getenv(env_key.c_str());
-  return (val && std::atoi(val) != 0) ? 1 : 0;
+  static bool cached = []() {
+    std::string env_key = "VLLM_KERNEL_OVERRIDE_BATCH_INVARIANT";
+    const char* val = std::getenv(env_key.c_str());
+    return (val && std::atoi(val) != 0) ? 1 : 0;
+  }();
+  return cached;
 }

 }  // namespace vllm
@@ -12,6 +12,7 @@ using CubMaxOp = cub::Max;
  #endif  // CUB_VERSION
 #else
  #include <hipcub/hipcub.hpp>
-using CubAddOp = cub::Sum;
-using CubMaxOp = cub::Max;
+namespace cub = hipcub;
+using CubAddOp = hipcub::Sum;
+using CubMaxOp = hipcub::Max;
 #endif  // USE_ROCM
@@ -2,7 +2,6 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

 import enum
-from typing import Union

 from cutlass_library import *

@@ -22,7 +21,7 @@ class MixedInputKernelScheduleType(enum.Enum):
    TmaWarpSpecializedCooperative = enum_auto()


-VLLMDataTypeNames: dict[Union[VLLMDataType, DataType], str] = {
+VLLMDataTypeNames: dict[VLLMDataType | DataType, str] = {
    **DataTypeNames,  # type: ignore
    **{
        VLLMDataType.u4b8: "u4b8",
@@ -30,7 +29,7 @@ VLLMDataTypeNames: dict[Union[VLLMDataType, DataType], str] = {
    },
 }

-VLLMDataTypeTag: dict[Union[VLLMDataType, DataType], str] = {
+VLLMDataTypeTag: dict[VLLMDataType | DataType, str] = {
    **DataTypeTag,  # type: ignore
    **{
        VLLMDataType.u4b8: "cutlass::vllm_uint4b8_t",
@@ -38,7 +37,7 @@ VLLMDataTypeTag: dict[Union[VLLMDataType, DataType], str] = {
    },
 }

-VLLMDataTypeSize: dict[Union[VLLMDataType, DataType], int] = {
+VLLMDataTypeSize: dict[VLLMDataType | DataType, int] = {
    **DataTypeSize,  # type: ignore
    **{
        VLLMDataType.u4b8: 4,
@@ -46,7 +45,7 @@ VLLMDataTypeSize: dict[Union[VLLMDataType, DataType], int] = {
    },
 }

-VLLMDataTypeVLLMScalarTypeTag: dict[Union[VLLMDataType, DataType], str] = {
+VLLMDataTypeVLLMScalarTypeTag: dict[VLLMDataType | DataType, str] = {
    VLLMDataType.u4b8: "vllm::kU4B8",
    VLLMDataType.u8b128: "vllm::kU8B128",
    DataType.u4: "vllm::kU4",
@@ -57,7 +56,7 @@ VLLMDataTypeVLLMScalarTypeTag: dict[Union[VLLMDataType, DataType], str] = {
    DataType.bf16: "vllm::kBfloat16",
 }

-VLLMDataTypeTorchDataTypeTag: dict[Union[VLLMDataType, DataType], str] = {
+VLLMDataTypeTorchDataTypeTag: dict[VLLMDataType | DataType, str] = {
    DataType.u8: "at::ScalarType::Byte",
    DataType.s8: "at::ScalarType::Char",
    DataType.e4m3: "at::ScalarType::Float8_e4m3fn",
@@ -67,9 +66,7 @@ VLLMDataTypeTorchDataTypeTag: dict[Union[VLLMDataType, DataType], str] = {
    DataType.f32: "at::ScalarType::Float",
 }

-VLLMKernelScheduleTag: dict[
-    Union[MixedInputKernelScheduleType, KernelScheduleType], str
-] = {
+VLLMKernelScheduleTag: dict[MixedInputKernelScheduleType | KernelScheduleType, str] = {
    **KernelScheduleTag,  # type: ignore
    **{
        MixedInputKernelScheduleType.TmaWarpSpecialized: "cutlass::gemm::KernelTmaWarpSpecialized",  # noqa: E501
@@ -6,7 +6,7 @@
 */

 #include "type_convert.cuh"
-#include "quantization/fp8/common.cuh"
+#include "quantization/w8a8/fp8/common.cuh"
 #include "dispatch_utils.h"
 #include "cub_helpers.h"
 #include "core/batch_invariant.hpp"
@@ -100,6 +100,11 @@ void apply_repetition_penalties_(torch::Tensor& logits,
                                 const torch::Tensor& output_mask,
                                 const torch::Tensor& repetition_penalties);

+void top_k_per_row(const torch::Tensor& logits, const torch::Tensor& rowStarts,
+                   const torch::Tensor& rowEnds, torch::Tensor& indices,
+                   torch::Tensor& values, int64_t numRows, int64_t stride0,
+                   int64_t stride1);
+
 void rms_norm_static_fp8_quant(torch::Tensor& out, torch::Tensor& input,
                               torch::Tensor& weight, torch::Tensor& scale,
                               double epsilon);
@@ -133,12 +138,12 @@ void silu_and_mul_nvfp4_quant(torch::Tensor& out,
                              torch::Tensor& input,
                              torch::Tensor& input_global_scale);
 #endif
-void silu_mul_fp8_quant_deep_gemm_cuda(
+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]
-    int64_t group_size, bool use_ue8m0, int64_t num_parallel_tokens);
+    bool use_ue8m0);

 void mul_and_silu(torch::Tensor& out, torch::Tensor& input);

@@ -7,7 +7,7 @@
 #include "../cuda_compat.h"
 #include "dispatch_utils.h"

-#include "quantization/fp8/common.cuh"
+#include "quantization/w8a8/fp8/common.cuh"

 #include <c10/util/Float8_e4m3fn.h>

@@ -114,13 +114,22 @@ __global__ void act_and_mul_quant_kernel(
 }

 __device__ __forceinline__ float silu(float x) {
-  return (__fdividef(x, (1.f + expf(-x))));
+  return __fdividef(x, (1.f + expf(-x)));
 }

 __device__ __forceinline__ float2 silu2(float2 x) {
  return make_float2(silu(x.x), silu(x.y));
 }

+__device__ __forceinline__ __nv_bfloat162 silu2_v2(float2 x) {
+#ifndef USE_ROCM
+  return make_bfloat162(__float2bfloat16_rn(silu(x.x)),
+                        __float2bfloat16_rn(silu(x.y)));
+#else
+  return __float22bfloat162_rn(make_float2(silu(x.x), silu(x.y)));
+#endif
+}
+
 #ifndef USE_ROCM
 __device__ __forceinline__ float warp_max(float v) {
  static constexpr unsigned FULL_MASK = 0xffffffffu;
@@ -223,224 +232,308 @@ constexpr __nv_bfloat16 get_fp8_min() {
    return __nv_bfloat16(__nv_bfloat16_raw{.x = 50032});
  }
 }
-#ifndef USE_ROCM
-template <typename fp8_type, int32_t NUM_WARPS, typename Idx_t,
-          int NUM_PARALLEL_TOKENS, bool USE_UE8M0, int GROUP_SIZE = 128,
+
+template <typename Idx_t>
+__device__ __forceinline__ int warp_expert_search(
+    int idx, int n, const Idx_t* __restrict__ input, Idx_t val) {
+  const Idx_t* input_ptr = input + idx;
+  int base_offset = 0;
+
+  for (;;) {
+    bool move_on = (idx < n && *input_ptr <= val);
+
+    unsigned mask = __ballot_sync(0xffffffff, move_on);
+
+    if (mask != 0xffffffffu) {
+      int last_lane = 31 - __clz(mask);
+      return base_offset + last_lane;
+    }
+
+    input_ptr += 32;
+    base_offset += 32;
+    idx += 32;
+  }
+}
+
+template <int num_parallel_tokens>
+__device__ __forceinline__ void token_bounds(int32_t n_tokens,
+                                             int32_t worker_id,
+                                             int32_t& n_tokens_lower,
+                                             int32_t& n_tokens_upper) {
+  if (n_tokens < num_parallel_tokens && worker_id < n_tokens) {
+    if (worker_id >= num_parallel_tokens) return;
+    n_tokens_lower = worker_id;
+    n_tokens_upper = worker_id + 1;
+  } else {
+    int32_t chunk_size = n_tokens / num_parallel_tokens;
+    int32_t residual = n_tokens - chunk_size * num_parallel_tokens;
+    auto calc_id = [&](int32_t id) {
+      if (id < residual)
+        return min(n_tokens, id * (chunk_size + 1));
+      else
+        return min(n_tokens, id * chunk_size + residual);
+    };
+    n_tokens_lower = calc_id(worker_id);
+    n_tokens_upper = calc_id(worker_id + 1);
+  }
+}
+
+template <int BLOCK_COUNT, int SMEM_SIZE_BYTES_Y, typename fp8_type,
+          int THREADS, typename Idx_t, bool USE_UE8M0, int GROUP_SIZE = 128,
          int NUM_STAGES = 3>
 __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
    const __nv_bfloat16* __restrict__ _input, fp8_type* __restrict__ _y_q,
-    float* __restrict__ _y_s, const int32_t* __restrict__ counts,
-
+    float* __restrict__ _y_s, const int32_t* __restrict__ tokens_per_expert,
    // sizes
-    int H, int G,
-
+    Idx_t E, Idx_t T, Idx_t H,
    // strides (in elements)
    Idx_t stride_i_e, Idx_t stride_i_t, Idx_t stride_i_h, Idx_t stride_yq_e,
    Idx_t stride_yq_t, Idx_t stride_yq_h, Idx_t stride_ys_e, Idx_t stride_ys_t,
    Idx_t stride_ys_g, Idx_t stride_counts_e) {
+#ifndef USE_ROCM
+  static constexpr int NUM_WARPS = THREADS / WARP_SIZE;
+
+  static constexpr int LOAD_STAGE_SIZE = 2 * GROUP_SIZE / 8;
+  static constexpr int LOAD_STAGE_MOD = NUM_STAGES * LOAD_STAGE_SIZE;
+
+  static constexpr int COMPUTE_STAGE_SIZE = 2 * GROUP_SIZE / 4;
+  static constexpr int COMPUTE_STAGE_MOD = COMPUTE_STAGE_SIZE * NUM_STAGES;
+
+  extern __shared__ __align__(16) __int128_t smem_128[];
+
+  int* s_expert_offsets =
+      reinterpret_cast<int*>(smem_128 + (SMEM_SIZE_BYTES_Y / 16));
+
  static constexpr __nv_bfloat16 fp8_min = get_fp8_min<fp8_type>();
  static constexpr __nv_bfloat16 fp8_max = get_fp8_max<fp8_type>();
-  // We assign EPS with its 16-bit unsigned counterpart to allow constexpr.
+  // We assign EPS with it's 16-bit unsigned counterpart to allow constexpr.
  static constexpr __nv_bfloat16 EPS = (__nv_bfloat16_raw{.x = 11996});
+  int tid = threadIdx.x;
+  int warp_id = tid >> 5;
+  int lane_id = tid & 0x1f;

-  // We pack 8 16-bit bfloat16 values into a 128-bit __int128_t.
-  static constexpr int32_t BFLOAT16_PER_GROUP = 8;
+  int running_sum{};
+  if (!warp_id) {
+    for (int i = 0; i < E; i += WARP_SIZE) {
+      bool valid = (i + threadIdx.x) < E;
+      int value =
+          (valid ? tokens_per_expert[i + threadIdx.x * stride_counts_e] : 0) +
+          (!lane_id ? running_sum : 0);

-  // We split the shared memory in half, corresponding to gate and up matrices:
-  // [...gate_i, ...up_i]  where 0 <= i < stages.
-  static constexpr int32_t S_NUM_128 =
-      2u * (GROUP_SIZE / BFLOAT16_PER_GROUP) * NUM_WARPS * NUM_STAGES;
-  static constexpr auto THREAD_COUNT = NUM_WARPS * WARP_SIZE;
-  static constexpr int HALF_THREAD_COUNT = THREAD_COUNT / 2;
-  static constexpr int32_t S_NUM_64 = S_NUM_128 * 2;
-  __shared__ __int128_t __align__(16) s_buff_128[S_NUM_128];
+      for (int offset = 1; offset < 32; offset *= 2) {
+        int n = __shfl_up_sync(0xFFFFFFFFu, value, offset);
+        if (lane_id >= offset) value += n;
+      }

-  const int32_t tid = threadIdx.x;
-  const int32_t warp_id = tid / WARP_SIZE;
-  const int32_t lane_id = tid % WARP_SIZE;
+      if (valid) {
+        s_expert_offsets[i + threadIdx.x + 1] = value;
+      }

-  auto s_buff_compute_32 = reinterpret_cast<__nv_bfloat162*>(s_buff_128);
+      running_sum = __shfl_sync(0xFFFFFFFFu, value, WARP_SIZE - 1);
+    }

-  // block handles one (expert e, group g)
-  int32_t pid = blockIdx.x;
-  int32_t e = pid / G;
-  int32_t g = pid % G;
-
-  const int32_t n_tokens = counts[e * stride_counts_e];
-
-  if (!n_tokens) {
-    return;  // Exit ASAP.
+    if (!lane_id) {
+      s_expert_offsets[0] = 0;
+    }
  }

-  const Idx_t stride_i_t_128 = stride_i_t / 8u;
+  __syncthreads();

-  int32_t n_tokens_lower, n_tokens_upper;
+  int32_t total_tokens = s_expert_offsets[E];

+  const int warp_position_yq = warp_id * (H / NUM_WARPS);
+  const int warp_position_scales = warp_id * (H / (GROUP_SIZE * NUM_WARPS));
+
+  // A single block will handle tokens_per_block tokens.
  // Each block i iterates over tokens of a slice of n_tokens =
  // expert_counts[i], with the size of chunk being
  // (n_tokens / NUM_PARALLEL_TOKENS) + residual, instead of
  // updiv(n_tokens, NUM_PARALLEL_TOKENS) for better scheduling.
-  if (n_tokens < NUM_PARALLEL_TOKENS && blockIdx.y < n_tokens) {
-    // Specialize this, but can be likely fused.
-    if (blockIdx.y >= NUM_PARALLEL_TOKENS) {
-      return;
-    }
-    n_tokens_lower = blockIdx.y;
-    n_tokens_upper = blockIdx.y + 1;
-  } else {
-    auto chunk_size = n_tokens / NUM_PARALLEL_TOKENS;
-    auto residual = n_tokens - chunk_size * NUM_PARALLEL_TOKENS;
-    auto calc_id = [&](int32_t id) {
-      if (id < residual) {
-        return min(n_tokens, id * (chunk_size + 1));
-      } else {
-        return min(n_tokens, id * chunk_size + residual);
-      }
-    };
-    n_tokens_lower = calc_id(blockIdx.y);
-    n_tokens_upper = calc_id(blockIdx.y + 1);
-  }

-  if (n_tokens_lower >= n_tokens_upper) {
+  // Each warp will get space to store its hidden dim for gate and up.
+  __int128_t* s_hidden_load = smem_128 + warp_id * ((2 * 128 / 8) * NUM_STAGES);
+  __int128_t* smem_load_ptr = s_hidden_load + lane_id;
+
+  const __nv_bfloat16 fp8_inv = __hdiv(__float2bfloat16(1.f), fp8_max);
+
+  int32_t compute_pipeline_offset_64 = 0;
+  int32_t load_stage_offset{};
+  const __nv_bfloat16 one_bf16 = __float2bfloat16_rn(1.f);
+
+  __int64_t* smem_compute_ptr = reinterpret_cast<__int64_t*>(smem_128) +
+                                warp_id * (2 * (GROUP_SIZE / 4) * NUM_STAGES) +
+                                lane_id;
+  __int64_t* s_gate64_ptr = smem_compute_ptr;
+  __int64_t* s_up64_ptr = smem_compute_ptr + GROUP_SIZE / 4;
+
+  int tokens_lower, tokens_upper;
+
+  token_bounds<BLOCK_COUNT>(total_tokens, blockIdx.x, tokens_lower,
+                            tokens_upper);
+
+  Idx_t expert_id{}, expert_offset{}, next_expert_offset{};
+  int token_id = tokens_lower;
+  int32_t t_load{};
+
+  if (token_id < tokens_upper) {
+    expert_id = warp_expert_search<int>(lane_id, E, s_expert_offsets, token_id);
+    expert_offset = s_expert_offsets[expert_id];
+    next_expert_offset = s_expert_offsets[expert_id + 1];
+  } else {
+    // This thread block has no work to do.
    return;
  }

-  // We do calculations here, using constexpr wherever possible.
-  const Idx_t base_i = e * stride_i_e + NUM_WARPS * g * GROUP_SIZE * stride_i_h;
-  const Idx_t base_ys = e * stride_ys_e + NUM_WARPS * g * stride_ys_g;
-  const Idx_t base_yq =
-      e * stride_yq_e + NUM_WARPS * g * GROUP_SIZE * stride_yq_h;
-  Idx_t gate_off_128 = (base_i / static_cast<Idx_t>(8u));
-  auto input_128_ptr = reinterpret_cast<const __int128_t*>(_input);
-  auto gate_128_ptr = input_128_ptr + gate_off_128 + (tid % HALF_THREAD_COUNT) +
-                      stride_i_t_128 * n_tokens_lower;
-  auto up_128_ptr = gate_128_ptr + (H * stride_i_h) / 8u;
-  auto y_s_ptr =
-      _y_s + base_ys + warp_id * stride_ys_g + n_tokens_lower * stride_ys_t;
-  auto y_q_ptr = _y_q + base_yq + warp_id * GROUP_SIZE +
-                 stride_yq_t * n_tokens_lower + 4 * lane_id;
-  int32_t t_load = n_tokens_lower, load_stage_id = 0;
-  auto s_buff_gate_load_128 = s_buff_128 + (tid % HALF_THREAD_COUNT);
-  auto s_buff_up_load_128 = s_buff_gate_load_128 + S_NUM_128 / 2u;
-  int32_t stage_offset{};
+  int t_load_bound = H / (GROUP_SIZE * NUM_WARPS);

-  static constexpr int32_t LOAD_STAGE_SIZE = (NUM_WARPS * WARP_SIZE / 2);
-  static constexpr int32_t LOAD_STAGE_MOD =
-      NUM_STAGES * (NUM_WARPS * WARP_SIZE / 2);
+  Idx_t base_i = ((expert_id * stride_i_e) / 8) +
+                 (token_id - expert_offset) * stride_i_t / 8;
+  const Idx_t gate_warp_offset =
+      warp_id * ((stride_i_h * H) / (8 * NUM_WARPS)) + (lane_id & 0b1111);
+
+  const __int128_t* input_128_ptr =
+      reinterpret_cast<const __int128_t*>(_input) + gate_warp_offset +
+      ((lane_id < 16) ? 0 : ((H * stride_i_h) / 8));
+  __int128_t* load_ptr = const_cast<__int128_t*>(input_128_ptr + base_i);
+
+  auto token_offset = token_id - expert_offset;

-  // Two halves of all threads in a block conduct global loads for gate and up,
-  // repsectively.
  auto load_and_advance_y_pred = [&] {
-    if (t_load < n_tokens_upper) {
-      auto s_gate_stage_128_staged_ptr = s_buff_gate_load_128 + stage_offset;
-      auto s_up_stage_128_staged_ptr = s_buff_up_load_128 + stage_offset;
+    if (t_load < t_load_bound) {
+      // Here we are simply continuing to load data
+      // from the current token.
+      auto smem_load_ptr_staged = smem_load_ptr + load_stage_offset;

      // It is very important that LOAD_STAGE_SIZE is constexpr to avoid
      // unnecessary ALU ops.
-      stage_offset += LOAD_STAGE_SIZE;
-      stage_offset %= LOAD_STAGE_MOD;
+      load_stage_offset += LOAD_STAGE_SIZE;
+      load_stage_offset %= LOAD_STAGE_MOD;

-      if (tid < HALF_THREAD_COUNT) {
-        cp_async4(s_gate_stage_128_staged_ptr, gate_128_ptr);
-        gate_128_ptr += stride_i_t_128;
+      cp_async4(smem_load_ptr_staged, load_ptr);
+      load_ptr += GROUP_SIZE / 8;
+      ++t_load;
+    } else if (token_id + 1 < tokens_upper) {
+      // We loaded everything from the current token, let's move on
+      // to the next one, and we checked that we have more tokens to load.
+      ++token_id;
+      t_load = 0;
+      if (token_id >= next_expert_offset) {
+        // We need to find the next expert.
+        do {
+          // This is a loop because it's possible
+          // that some experts are assigned 0 tokens.
+          // NOTE: We are guaranteed that there's at least
+          // one more token left so we don't have to check for
+          // expert_id bounds.
+          ++expert_id;
+          // This skips 1 memory read.
+          expert_offset = next_expert_offset;
+          next_expert_offset = s_expert_offsets[expert_id + 1];
+        } while (next_expert_offset == expert_offset);
+
+        base_i = expert_id * (stride_i_e / 8);
+        token_offset = 0;
+        load_ptr = const_cast<__int128_t*>(input_128_ptr + base_i);
      } else {
-        cp_async4(s_up_stage_128_staged_ptr, up_128_ptr);
-        up_128_ptr += stride_i_t_128;
+        // We remain within the same expert, so just
+        // move by H/4 __int128_t (2 * H/8).
+        base_i += stride_yq_t / 4;
+        token_offset++;
      }
+
+      load_ptr = const_cast<__int128_t*>(input_128_ptr + base_i);
+
+      auto smem_load_ptr_staged = smem_load_ptr + load_stage_offset;
+
+      // It is very important that LOAD_STAGE_SIZE is constexpr to avoid
+      // unnecessary ALU ops.
+      load_stage_offset += LOAD_STAGE_SIZE;
+      load_stage_offset %= LOAD_STAGE_MOD;
+
+      cp_async4(smem_load_ptr_staged, load_ptr);
+      load_ptr += GROUP_SIZE / 8;
      ++t_load;
-      ++load_stage_id;
    }
    // We fence even if there is nothing to load to simplify pipelining.
    cp_async_fence();
  };

+  // We need to warm-up the pipeline.
  #pragma unroll
  for (int i = 0; i < NUM_STAGES - 1; i++) {
    load_and_advance_y_pred();
  }

-  __int64_t* s_gate_ptr = reinterpret_cast<__int64_t*>(
-                              s_buff_compute_32 + warp_id * (GROUP_SIZE / 2)) +
-                          lane_id;
-  __int64_t* s_up_ptr = s_gate_ptr + S_NUM_64 / 2;
+  __nv_fp8x4_e4m3* y_q_base_ptr =
+      reinterpret_cast<__nv_fp8x4_e4m3*>(_y_q) + lane_id;
+  auto y_scale_base_ptr = _y_s + warp_position_scales * stride_ys_g;

-  static constexpr int32_t STAGE_SIZE = (GROUP_SIZE * NUM_WARPS) / 4u;
-  static constexpr int32_t STAGE_MOD = STAGE_SIZE * NUM_STAGES;
+  for (auto j = tokens_lower; j < tokens_upper; j++) {
+    const Idx_t base_ys = expert_id * stride_ys_e;
+    auto y_s_ptr = y_scale_base_ptr + base_ys + token_offset * stride_ys_t;
+    __nv_fp8x4_e4m3* y_q_ptr =
+        y_q_base_ptr + (expert_id * stride_yq_e + token_offset * stride_yq_t +
+                        warp_position_yq * stride_yq_h) /
+                           4;
+    const int COMPUTE_LIMIT = H / (GROUP_SIZE * NUM_WARPS);

-  int32_t compute_pipeline_offset_64 = 0;
+    for (int i = 0; i < COMPUTE_LIMIT; i++) {
+      cp_async_wait<NUM_STAGES - 2>();
+      __syncthreads();
+      load_and_advance_y_pred();

-  for (int32_t t = n_tokens_lower; t < n_tokens_upper; ++t) {
-    __nv_bfloat162 results_bf162[2];
+      __int64_t* gate64_ptr = s_gate64_ptr + compute_pipeline_offset_64;
+      __int64_t* up64_ptr = s_up64_ptr + compute_pipeline_offset_64;

-    cp_async_wait<NUM_STAGES - 2>();
-    __syncthreads();
+      // COMPUTE_STAGE_SIZE/MOD must also be constexpr!
+      compute_pipeline_offset_64 += COMPUTE_STAGE_SIZE;
+      compute_pipeline_offset_64 %= COMPUTE_STAGE_MOD;

-    // We double-buffer pipelined loads so that the next load will
-    // concurrently run with compute without overwrites.
-    load_and_advance_y_pred();
+      __int64_t gate64 = *gate64_ptr;
+      __int64_t up64 = *up64_ptr;

-    auto s_gate_compute_64 = s_gate_ptr + compute_pipeline_offset_64;
-    auto s_up_compute_64 = s_up_ptr + compute_pipeline_offset_64;
-
-    // STAGE_SIZE must also be constexpr!
-    compute_pipeline_offset_64 += STAGE_SIZE;
-    compute_pipeline_offset_64 %= STAGE_MOD;
-
-    // Each thread loads (gate/up) 2X 4X bfloat16 values into registers.
-    __int64_t gate64 = *s_gate_compute_64;
-    __nv_bfloat162* s_gate_compute_32 =
-        reinterpret_cast<__nv_bfloat162*>(&gate64);
-
-    __int64_t up64 = *s_up_compute_64;
-    __nv_bfloat162* s_up_compute_32 = reinterpret_cast<__nv_bfloat162*>(&up64);
+      // Compute
+      __nv_bfloat162 res[2];
+      __nv_bfloat162* s_up_comp = reinterpret_cast<__nv_bfloat162*>(&up64);
+      __nv_bfloat162* s_gate_comp = reinterpret_cast<__nv_bfloat162*>(&gate64);

  #pragma unroll
-    for (int i = 0; i < 2; i++) {
-      // For silu, we make sure that div is emitted.
-      float2 gate = silu2(__bfloat1622float2(s_gate_compute_32[i]));
-      results_bf162[i] = __float22bfloat162_rn(gate);
-    }
+      for (int32_t k = 0; k < 2; ++k) {
+        __nv_bfloat162 gate = silu2_v2(__bfloat1622float2(s_gate_comp[k]));
+        res[k] = __hmul2(gate, s_up_comp[k]);
+      }
+
+      auto _y_max2 = __hmax2(__habs2(res[0]), __habs2(res[1]));
+
+      _y_max2.x = __hmax(__hmax(_y_max2.x, _y_max2.y), EPS);
+
+      __nv_bfloat16 y_s = __hmul(warp_max(_y_max2.x), fp8_inv);
+
+      if constexpr (USE_UE8M0) {
+        y_s = hexp2(hceil(hlog2(y_s)));
+      }
+
+      __nv_bfloat16 inv_y = __hdiv(one_bf16, y_s);
+
+      auto y_s2 = make_bfloat162(inv_y, inv_y);

  #pragma unroll
-    for (int i = 0; i < 2; i++) {
-      results_bf162[i] = __hmul2(results_bf162[i], s_up_compute_32[i]);
-    }
+      for (int32_t k = 0; k < 2; ++k) {
+        res[k] = clip(__hmul2(res[k], y_s2), __bfloat162bfloat162(fp8_min),
+                      __bfloat162bfloat162(fp8_max));
+      }

-    auto _y_max2 =
-        __hmax2(__habs2(results_bf162[0]), __habs2(results_bf162[1]));
+      *y_q_ptr = __nv_fp8x4_e4m3(res[0], res[1]);
+      y_q_ptr += WARP_SIZE * stride_yq_h;

-    __nv_bfloat16 y_max_bf16 = __hmax(EPS, __hmax(_y_max2.x, _y_max2.y));
-
-    // An entire group is assigned to a single warp, so a simple warp reduce
-    // is used.
-    __nv_bfloat16 y_s = warp_max(y_max_bf16) / fp8_max;
-
-    if constexpr (USE_UE8M0) {
-      y_s = hexp2(hceil(hlog2(y_s)));
-    }
-
-    auto inv_y = __float2bfloat16_rn(1.f) / y_s;
-
-    auto y_s2 = make_bfloat162(inv_y, inv_y);
-
-  #pragma unroll
-    for (int32_t i = 0; i < 2; ++i) {
-      results_bf162[i] =
-          clip(__hmul2(results_bf162[i], y_s2), __bfloat162bfloat162(fp8_min),
-               __bfloat162bfloat162(fp8_max));
-    }
-
-    auto fp8x4 = __nv_fp8x4_e4m3(results_bf162[0], results_bf162[1]);
-    *reinterpret_cast<__nv_fp8x4_e4m3*>(y_q_ptr) = fp8x4;
-    y_q_ptr += stride_yq_t;
-
-    if (lane_id == 0) {
-      *y_s_ptr = y_s;
-      y_s_ptr += stride_ys_t;
+      if (!lane_id) {
+        *y_s_ptr = y_s;
+        y_s_ptr += stride_ys_g;
+      }
    }
  }
-}
 #endif
+}

 }  // namespace vllm

@@ -475,14 +568,14 @@ void silu_and_mul_quant(torch::Tensor& out,    // [..., d]
  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel);
 }

-void silu_mul_fp8_quant_deep_gemm_cuda(
-    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]
-    int64_t group_size, bool use_ue8m0, int64_t num_parallel_tokens) {
+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]
+    bool use_ue8m0) {
 #ifndef USE_ROCM
-  // This kernel relies heavily on cp.async and fp8 support.
+
  // This kernel currently only supports H % 128 == 0 and assumes a
  // fixed GROUP_SIZE of 128.
  TORCH_CHECK(input.dtype() == torch::kBFloat16);
@@ -491,10 +584,6 @@ void silu_mul_fp8_quant_deep_gemm_cuda(
  TORCH_CHECK(y_s.dtype() == torch::kFloat32);
  TORCH_CHECK(input.size(-1) % 256 == 0);

-  // Check that num_parallel_tokens is of power of 2 and between 1 and 64.
-  TORCH_CHECK(1 <= num_parallel_tokens && num_parallel_tokens <= 64);
-  TORCH_CHECK(!(num_parallel_tokens & (num_parallel_tokens - 1)));
-
  using Idx_t = int64_t;

  Idx_t E = input.size(0);
@@ -510,81 +599,54 @@ void silu_mul_fp8_quant_deep_gemm_cuda(
  Idx_t stride_ys_t = y_s.stride(1);
  Idx_t stride_ys_g = y_s.stride(2);

-  Idx_t stride_counts_e = counts.stride(0);
+  Idx_t stride_counts_e = tokens_per_expert.stride(0);

  static constexpr int GROUP_SIZE = 128;

-  #define KERNEL_FN                                                         \
-    if (use_ue8m0) {                                                        \
-      vllm::silu_mul_fp8_quant_deep_gemm_kernel<fp8_t, NUM_WARPS, Idx_t,    \
-                                                NUM_PARALLEL_TOKENS, true>  \
-          <<<grid, block, 0, stream>>>(                                     \
-              reinterpret_cast<__nv_bfloat16*>(input.data_ptr()),           \
-              (fp8_t*)y_q.data_ptr(), y_s.data_ptr<float>(),                \
-              reinterpret_cast<int32_t*>(counts.data_ptr<int>()), H, G,     \
-              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_counts_e);                                             \
-    } else {                                                                \
-      vllm::silu_mul_fp8_quant_deep_gemm_kernel<fp8_t, NUM_WARPS, Idx_t,    \
-                                                NUM_PARALLEL_TOKENS, false> \
-          <<<grid, block, 0, stream>>>(                                     \
-              reinterpret_cast<__nv_bfloat16*>(input.data_ptr()),           \
-              (fp8_t*)y_q.data_ptr(), y_s.data_ptr<float>(),                \
-              reinterpret_cast<int32_t*>(counts.data_ptr<int>()), H, G,     \
-              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_counts_e);                                             \
-    }
-
-  #define KERNEL_CALL_H                                       \
-    if (H % (4 * GROUP_SIZE) == 0) {                          \
-      static constexpr int NUM_WARPS = 4;                     \
-      populate_launch_params(NUM_WARPS, NUM_PARALLEL_TOKENS); \
-      KERNEL_FN                                               \
-    } else {                                                  \
-      static constexpr int NUM_WARPS = 1;                     \
-      populate_launch_params(NUM_WARPS, NUM_PARALLEL_TOKENS); \
-      KERNEL_FN                                               \
-    }
-
-  #define KERNEL_CALL_TOP_LEVEL                      \
-    if (num_parallel_tokens == 1) {                  \
-      static constexpr int NUM_PARALLEL_TOKENS = 1;  \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 2) {           \
-      static constexpr int NUM_PARALLEL_TOKENS = 2;  \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 4) {           \
-      static constexpr int NUM_PARALLEL_TOKENS = 4;  \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 8) {           \
-      static constexpr int NUM_PARALLEL_TOKENS = 8;  \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 16) {          \
-      static constexpr int NUM_PARALLEL_TOKENS = 16; \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 32) {          \
-      static constexpr int NUM_PARALLEL_TOKENS = 32; \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 64) {          \
-      static constexpr int NUM_PARALLEL_TOKENS = 64; \
-      KERNEL_CALL_H                                  \
-    }
-
-  Idx_t G;
-  dim3 block, grid;
-  auto populate_launch_params = [&](int num_warps, int _num_parallel_tokens) {
-    G = H / Idx_t(group_size * num_warps);
-    grid = dim3(E * G, _num_parallel_tokens);
-    block = dim3(num_warps * WARP_SIZE);
-  };
-
  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
-  VLLM_DISPATCH_FP8_TYPES(y_q.scalar_type(),
-                          "silu_mul_fp8_quant_deep_gemm_kernel",
-                          [&] { KERNEL_CALL_TOP_LEVEL });
+
+  #define KERNEL(BLOCK_COUNT, USE_UE8M0, THREAD_COUNT, STAGES)                 \
+    static constexpr int NUM_WARPS = THREAD_COUNT / WARP_SIZE;                 \
+    int sms = SILU_V2_BLOCK_COUNT;                                             \
+    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(                                                   \
+        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, THREAD_COUNT, Idx_t,   \
+              USE_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(), y_s.data_ptr<float>(),               \
+                  reinterpret_cast<int32_t*>(tokens_per_expert.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_counts_e);                               \
+        });
+
+  static constexpr int SILU_V2_BLOCK_COUNT = 132 * 32;
+
+  if (!use_ue8m0) {
+    if (H >= 4096) {
+      static constexpr int NUM_STAGES = 4;
+      static constexpr int THREAD_COUNT = 256;
+      KERNEL(SILU_V2_BLOCK_COUNT, false, THREAD_COUNT, NUM_STAGES);
+    } else {
+      static constexpr int THREAD_COUNT = 32;
+      KERNEL(SILU_V2_BLOCK_COUNT, false, THREAD_COUNT, 2);
+    }
+  } else {
+    if (H >= 4096) {
+      static constexpr int NUM_STAGES = 4;
+      static constexpr int THREAD_COUNT = 256;
+      KERNEL(SILU_V2_BLOCK_COUNT, true, THREAD_COUNT, NUM_STAGES);
+    } else {
+      static constexpr int THREAD_COUNT = 32;
+      KERNEL(SILU_V2_BLOCK_COUNT, true, THREAD_COUNT, 2);
+    }
+  }

 #endif
 }
@@ -6,7 +6,7 @@

 #include "quantization/vectorization.cuh"
 // TODO(luka/varun):refactor common.cuh to use this file instead
-#include "quantization/fp8/common.cuh"
+#include "quantization/w8a8/fp8/common.cuh"

 namespace vllm {

@@ -9,7 +9,6 @@ from collections.abc import Iterable
 from copy import deepcopy
 from dataclasses import dataclass, fields
 from functools import reduce
-from typing import Optional, Union

 import jinja2
 from vllm_cutlass_library_extension import (
@@ -259,7 +258,7 @@ class ScheduleConfig:
@dataclass(frozen=True)
 class TypeConfig:
    a: DataType
-    b: Union[DataType, VLLMDataType]
+    b: DataType | VLLMDataType
    b_group_scale: DataType
    b_group_zeropoint: DataType
    b_channel_scale: DataType
@@ -280,7 +279,7 @@ class PrepackTypeConfig:
 class ImplConfig:
    types: TypeConfig
    schedules: list[ScheduleConfig]
-    heuristic: list[tuple[Optional[str], ScheduleConfig]]
+    heuristic: list[tuple[str | None, ScheduleConfig]]


 def generate_sch_sig(schedule_config: ScheduleConfig) -> str:
@@ -5,7 +5,7 @@
 #include <hip/hip_bf16.h>
 #include <hip/hip_bfloat16.h>

-#include "../../../attention/attention_dtypes.h"
+#include "../../../../attention/attention_dtypes.h"

 namespace vllm {
 #ifdef USE_ROCM
@@ -1,7 +1,7 @@
 #include "common.cuh"
 #include "dispatch_utils.h"
-#include "../../cub_helpers.h"
-#include "../vectorization_utils.cuh"
+#include "cub_helpers.h"
+#include "quantization/vectorization_utils.cuh"
 #include <c10/cuda/CUDAGuard.h>
 #include <ATen/cuda/Exceptions.h>

@@ -1,6 +1,6 @@
 #pragma once

-#include "../../../attention/attention_dtypes.h"
+#include "../../../../attention/attention_dtypes.h"
 #include <assert.h>
 #include <float.h>
 #include <stdint.h>
@@ -1,6 +1,6 @@
 #include <ATen/cuda/CUDAContext.h>

-#include "../per_token_group_quant_8bit.h"
+#include "quantization/w8a8/per_token_group_quant_8bit.h"

 #include <cmath>

@@ -8,9 +8,9 @@

 #include <torch/all.h>

-#include "../vectorization.cuh"
-#include "../vectorization_utils.cuh"
-#include "../../dispatch_utils.h"
+#include "quantization/vectorization.cuh"
+#include "quantization/vectorization_utils.cuh"
+#include "dispatch_utils.h"

 __device__ __forceinline__ float GroupReduceMax(float val) {
  unsigned mask = threadIdx.x % 32 >= 16 ? 0xffff0000 : 0x0000ffff;
@@ -212,4 +212,4 @@ void per_token_group_quant_fp8(const torch::Tensor& input,
                               double fp8_max, bool scale_ue8m0) {
  per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
                             fp8_min, fp8_max, scale_ue8m0);
-}
+}
@@ -0,0 +1,12 @@
+#include <ATen/cuda/CUDAContext.h>
+#include <torch/all.h>
+
+#include "quantization/w8a8/per_token_group_quant_8bit.h"
+
+void per_token_group_quant_int8(const torch::Tensor& input,
+                                torch::Tensor& output_q,
+                                torch::Tensor& output_s, int64_t group_size,
+                                double eps, double int8_min, double int8_max) {
+  per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
+                             int8_min, int8_max);
+}
--- a/Show More
+++ b/Show More