Remove executable flag on a few files

Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>

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

@@ -51,6 +51,7 @@ function cpu_tests() {
     pytest -v -s tests/kernels/attention/test_cache.py -m cpu_model
     pytest -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
     pytest -v -s tests/models/language/generation -m cpu_model
+    VLLM_CPU_SGL_KERNEL=1 pytest -v -s tests/models/language/generation -m cpu_model
     pytest -v -s tests/models/language/pooling -m cpu_model
     pytest -v -s tests/models/multimodal/generation \
                 --ignore=tests/models/multimodal/generation/test_mllama.py \
@@ -98,4 +99,4 @@ function cpu_tests() {
 
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 1h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 1.5h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"

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

@@ -2,10 +2,34 @@
 
 # This script build the CPU docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
-set -ex
+set -exuo pipefail
 
 # Try building the docker image
-docker build -t hpu-test-env -f docker/Dockerfile.hpu .
+cat <<EOF | docker build -t hpu-plugin-v1-test-env -f - .
+FROM 1.22-413-pt2.7.1:latest
+
+COPY ./ /workspace/vllm
+
+WORKDIR /workspace/vllm
+
+RUN pip install -v -r requirements/hpu.txt
+RUN pip install git+https://github.com/vllm-project/vllm-gaudi.git
+
+ENV no_proxy=localhost,127.0.0.1
+ENV PT_HPU_ENABLE_LAZY_COLLECTIVES=true
+
+RUN VLLM_TARGET_DEVICE=hpu python3 setup.py install
+
+# install development dependencies (for testing)
+RUN python3 -m pip install -e tests/vllm_test_utils
+
+WORKDIR /workspace/
+
+RUN git clone https://github.com/vllm-project/vllm-gaudi.git
+
+RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
+EOF
 
 # Setup cleanup
 # certain versions of HPU software stack have a bug that can
@@ -14,13 +38,21 @@ docker build -t hpu-test-env -f docker/Dockerfile.hpu .
 # functions, while other platforms only need one remove_docker_container
 # function.
 EXITCODE=1
-remove_docker_containers() { docker rm -f hpu-test || true; docker rm -f hpu-test-tp2 || true; }
-remove_docker_containers_and_exit() { remove_docker_containers; exit $EXITCODE; }
-trap remove_docker_containers_and_exit EXIT
+remove_docker_containers() { docker rm -f hpu-plugin-v1-test || true; }
+trap 'remove_docker_containers; exit $EXITCODE;' EXIT
 remove_docker_containers
 
-# Run the image and launch offline inference
-docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
-docker run --runtime=habana --name=hpu-test-tp2 --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --tensor-parallel-size 2
+echo "Running HPU plugin v1 test"
+docker run --rm --runtime=habana --name=hpu-plugin-v1-test --network=host \
+  -e HABANA_VISIBLE_DEVICES=all \
+  hpu-plugin-v1-test-env \
+  /bin/bash "/workspace/vllm-gaudi/tests/upstream_tests/ci_tests.sh"
 
 EXITCODE=$?
+if [ $EXITCODE -eq 0 ]; then
+  echo "Test with basic model passed"
+else
+  echo "Test with basic model FAILED with exit code: $EXITCODE" >&2
+fi
+
+# The trap will handle the container removal and final exit.

.buildkite/test-pipeline.yaml

@@ -536,6 +536,17 @@ steps:
     - pip freeze | grep -E 'torch'
     - pytest -v -s models/language -m core_model
 
+- label: Language Models Test (Hybrid) # 35 min
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/language/generation
+  commands:
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
+    - pytest -v -s models/language/generation -m hybrid_model
+
 - label: Language Models Test (Extended Generation) # 1hr20min
   mirror_hardwares: [amdexperimental]
   optional: true
@@ -545,7 +556,7 @@ steps:
   commands:
     # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
     - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
-    - pytest -v -s models/language/generation -m 'not core_model'
+    - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 - label: Language Models Test (Extended Pooling)  # 36min
   mirror_hardwares: [amdexperimental]

.github/mergify.yml

@@ -27,6 +27,22 @@ pull_request_rules:
       add:
         - ci/build
 
+- name: label-deepseek
+  description: Automatically apply deepseek label
+  conditions:
+    - or:
+      - files~=^examples/.*deepseek.*\.py
+      - files~=^tests/.*deepseek.*\.py
+      - files~=^vllm/entrypoints/openai/tool_parsers/.*deepseek.*\.py
+      - files~=^vllm/model_executor/models/.*deepseek.*\.py
+      - files~=^vllm/reasoning/.*deepseek.*\.py
+      - files~=^vllm/transformers_utils/.*deepseek.*\.py
+      - title~=(?i)DeepSeek
+  actions:
+    label:
+      add:
+        - deepseek
+
 - name: label-frontend
   description: Automatically apply frontend label
   conditions:

.pre-commit-config.yaml

@@ -160,6 +160,13 @@ repos:
     types: [python]
     pass_filenames: false
     additional_dependencies: [pathspec, regex]
+  - id: validate-config
+    name: Validate configuration has default values and that each field has a docstring
+    entry: python tools/validate_config.py
+    language: python
+    types: [python]
+    pass_filenames: true
+    files: vllm/config.py|tests/test_config.py
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

CMakeLists.txt

@@ -420,6 +420,36 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
+
+  # The cutlass_scaled_mm kernels for Geforce Blackwell SM120 (c3x, i.e. CUTLASS 3.x) require
+  # CUDA 12.8 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 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"
+    )
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM120=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm120 for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm120 as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
+                     "later if you intend on running FP8 quantized models on "
+                     "Blackwell.")
+    else()
+      message(STATUS "Not building scaled_mm_c3x_120 as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
+
+
   # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
   # require CUDA 12.8 or later
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
@@ -562,7 +592,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                      "if you intend on running FP8 quantized MoE models on Hopper.")
     else()
       message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
-                     "in CUDA target architectures")
+                     "in CUDA target architectures.")
     endif()
   endif()
 
@@ -574,7 +604,17 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       SRCS "${SRCS}"
       CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
-  endif() 
+    message(STATUS "Building moe_data for archs: ${CUTLASS_MOE_DATA_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
+      message(STATUS "Not building moe_data as CUDA Compiler version is "
+                     "not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
+                     "if you intend on running FP8 quantized MoE models on Hopper or Blackwell.")
+    else()
+      message(STATUS "Not building moe_data as no compatible archs found "
+                     "in CUDA target architectures.")
+    endif()
+  endif()
 
   #
   # Machete kernels

benchmarks/benchmark_serving.py

@@ -551,7 +551,7 @@ async def benchmark(
         "total_input_tokens": metrics.total_input,
         "total_output_tokens": metrics.total_output,
         "request_throughput": metrics.request_throughput,
-        "request_goodput:": metrics.request_goodput if goodput_config_dict else None,
+        "request_goodput": metrics.request_goodput if goodput_config_dict else None,
         "output_throughput": metrics.output_throughput,
         "total_token_throughput": metrics.total_token_throughput,
         "input_lens": [output.prompt_len for output in outputs],

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -113,6 +113,7 @@ def bench_run(
         w2_scale: torch.Tensor,
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
+        per_act_token: bool,
         num_repeats: int,
     ):
         for _ in range(num_repeats):
@@ -124,7 +125,8 @@ def bench_run(
                 topk_ids,
                 w1_scale,
                 w2_scale,
-                a1_scale=a_scale,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_cutlass_from_graph(
@@ -148,7 +150,8 @@ def bench_run(
                 topk_ids,
                 w1_scale,
                 w2_scale,
-                a1_scale=a_scale,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_triton_from_graph(
@@ -227,6 +230,7 @@ def bench_run(
         "w2_q": w2_q,
         "w1_scale": w1_scale,
         "w2_scale": w2_scale,
+        "per_act_token": per_act_token,
         # cuda graph params
         "cutlass_graph": cutlass_graph,
         "triton_graph": triton_graph,
@@ -287,12 +291,13 @@ def bench_run(
         w2_scale,
         topk_weights,
         topk_ids,
+        per_act_token,
         num_warmup,
     )
 
     results.append(
         benchmark.Timer(
-            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, num_runs)",  # noqa: E501
+            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, per_act_token, num_runs)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,

benchmarks/kernels/benchmark_machete.py

@@ -234,8 +234,10 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
 
         fn = lambda: ops.gptq_marlin_gemm(
             a=bt.a,
+            c=None,
             b_q_weight=w_q,
             b_scales=w_s,
+            global_scale=None,
             b_zeros=w_zp,
             g_idx=g_idx,
             perm=sort_indices,

cmake/cpu_extension.cmake

@@ -96,12 +96,21 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
             CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
             list(APPEND CXX_COMPILE_FLAGS "-mavx512bf16")
+            set(ENABLE_AVX512BF16 ON)
         else()
+            set(ENABLE_AVX512BF16 OFF)
             message(WARNING "Disable AVX512-BF16 ISA support, requires gcc/g++ >= 12.3")
         endif()
     else()
+        set(ENABLE_AVX512BF16 OFF)
         message(WARNING "Disable AVX512-BF16 ISA support, no avx512_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512BF16=1.")
     endif()
+
+    find_isa(${CPUINFO} "avx512_vnni" AVX512VNNI_FOUND)
+    if (AVX512VNNI_FOUND)
+        list(APPEND CXX_COMPILE_FLAGS "-mavx512vnni")
+        set(ENABLE_AVX512VNNI ON) 
+    endif() 
     
 elseif (AVX2_FOUND)
     list(APPEND CXX_COMPILE_FLAGS "-mavx2")
@@ -231,6 +240,17 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         "csrc/cpu/quant.cpp"
         "csrc/cpu/shm.cpp"
         ${VLLM_EXT_SRC})
+    if (ENABLE_AVX512BF16 AND ENABLE_AVX512VNNI)
+        set(VLLM_EXT_SRC
+            "csrc/cpu/sgl-kernels/gemm.cpp"
+            "csrc/cpu/sgl-kernels/gemm_int8.cpp"
+            "csrc/cpu/sgl-kernels/gemm_fp8.cpp"
+            "csrc/cpu/sgl-kernels/moe.cpp"
+            "csrc/cpu/sgl-kernels/moe_int8.cpp"
+            "csrc/cpu/sgl-kernels/moe_fp8.cpp"
+            ${VLLM_EXT_SRC})
+        add_compile_definitions(-DCPU_CAPABILITY_AVX512)
+    endif()
 elseif(POWER10_FOUND)
     set(VLLM_EXT_SRC
         "csrc/cpu/quant.cpp"

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 5f3644181c7a15345ce20bfc65af117d3601b524
+          GIT_TAG 1c2624e53c078854e0637ee566c72fe2107e75f4
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

@@ -265,8 +265,8 @@ macro(set_gencode_flags_for_srcs)
 endmacro()
 
 #
-# For the given `SRC_CUDA_ARCHS` list of gencode versions in the form 
-#  `<major>.<minor>[letter]` compute the "loose intersection" with the 
+# For the given `SRC_CUDA_ARCHS` list of gencode versions in the form
+#  `<major>.<minor>[letter]` compute the "loose intersection" with the
 #  `TGT_CUDA_ARCHS` list of gencodes. We also support the `+PTX` suffix in
 #  `SRC_CUDA_ARCHS` which indicates that the PTX code should be built when there
 #  is a CUDA_ARCH in `TGT_CUDA_ARCHS` that is equal to or larger than the
@@ -278,7 +278,7 @@ endmacro()
 #  in `SRC_CUDA_ARCHS` that is less or equal to the version in `TGT_CUDA_ARCHS`.
 # We have special handling for x.0a, if x.0a is in `SRC_CUDA_ARCHS` and x.0 is
 #  in `TGT_CUDA_ARCHS` then we should remove x.0a from `SRC_CUDA_ARCHS` and add
-#  x.0a to the result (and remove x.0 from TGT_CUDA_ARCHS). 
+#  x.0a to the result (and remove x.0 from TGT_CUDA_ARCHS).
 # The result is stored in `OUT_CUDA_ARCHS`.
 #
 # Example:
@@ -313,21 +313,16 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   # if x.0a is in SRC_CUDA_ARCHS and x.0 is in CUDA_ARCHS then we should
   # remove x.0a from SRC_CUDA_ARCHS and add x.0a to _CUDA_ARCHS
   set(_CUDA_ARCHS)
-  if ("9.0a" IN_LIST _SRC_CUDA_ARCHS)
-    list(REMOVE_ITEM _SRC_CUDA_ARCHS "9.0a")
-    if ("9.0" IN_LIST TGT_CUDA_ARCHS)
-      list(REMOVE_ITEM _TGT_CUDA_ARCHS "9.0")
-      set(_CUDA_ARCHS "9.0a")
+  foreach(_arch ${_SRC_CUDA_ARCHS})
+    if(_arch MATCHES "\\a$")
+      list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
+      string(REPLACE "a" "" _base "${_arch}")
+      if ("${_base}" IN_LIST TGT_CUDA_ARCHS)
+        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
+        list(APPEND _CUDA_ARCHS "${_arch}")
+      endif()
     endif()
-  endif()
-
-  if ("10.0a" IN_LIST _SRC_CUDA_ARCHS)
-    list(REMOVE_ITEM _SRC_CUDA_ARCHS "10.0a")
-    if ("10.0" IN_LIST TGT_CUDA_ARCHS)
-      list(REMOVE_ITEM _TGT_CUDA_ARCHS "10.0")
-      set(_CUDA_ARCHS "10.0a")
-    endif()
-  endif()
+  endforeach()
 
   list(SORT _SRC_CUDA_ARCHS COMPARE NATURAL ORDER ASCENDING)
 
@@ -359,7 +354,7 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   endforeach()
 
   list(REMOVE_DUPLICATES _CUDA_ARCHS)
-  
+
   # reapply +PTX suffix to architectures that requested PTX
   set(_FINAL_ARCHS)
   foreach(_arch ${_CUDA_ARCHS})
@@ -370,7 +365,7 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
     endif()
   endforeach()
   set(_CUDA_ARCHS ${_FINAL_ARCHS})
-  
+
   set(${OUT_CUDA_ARCHS} ${_CUDA_ARCHS} PARENT_SCOPE)
 endfunction()
 

csrc/cpu/sgl-kernels/common.h

@@ -0,0 +1,238 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#pragma once
+
+#include <ATen/ATen.h>
+#include <ATen/Parallel.h>
+#include <ATen/record_function.h>
+
+// clang-format off
+
+#if defined(_OPENMP)
+#include <omp.h>
+#endif
+
+namespace {
+
+// dispatch bool
+#define AT_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...)                                 \
+  [&] {                                                                          \
+    if (BOOL_V) {                                                                \
+      constexpr bool BOOL_NAME = true;                                           \
+      return __VA_ARGS__();                                                      \
+    } else {                                                                     \
+      constexpr bool BOOL_NAME = false;                                          \
+      return __VA_ARGS__();                                                      \
+    }                                                                            \
+  }()
+
+// dispatch: bfloat16, float16, int8_t, fp8_e4m3
+#define CPU_DISPATCH_PACKED_TYPES(TYPE, ...)                                    \
+  [&] {                                                                         \
+    switch (TYPE) {                                                             \
+      case at::ScalarType::BFloat16 : {                                         \
+        using packed_t = at::BFloat16;                                          \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      case at::ScalarType::Half: {                                              \
+        using packed_t = at::Half;                                              \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      case at::ScalarType::Char : {                                             \
+        using packed_t = int8_t;                                                \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      case at::ScalarType::Float8_e4m3fn : {                                    \
+        using packed_t = at::Float8_e4m3fn;                                     \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      default:                                                                  \
+        TORCH_CHECK(false, "Unsupported floating data type.\n");                \
+    }                                                                           \
+  }()
+
+#define UNUSED(x) (void)(x)
+
+#define CHECK_CPU(x) TORCH_CHECK(x.device().type() == at::kCPU, #x " must be a CPU tensor")
+
+#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_LAST_DIM_CONTIGUOUS(x) \
+  TORCH_CHECK(x.strides()[x.strides().size() - 1] == 1, #x "must be contiguous at last dimention")
+
+#define CHECK_INPUT(x) \
+  CHECK_CPU(x);        \
+  CHECK_CONTIGUOUS(x)
+#define CHECK_LAST_DIM_CONTIGUOUS_INPUT(x) \
+  CHECK_CPU(x);                            \
+  CHECK_LAST_DIM_CONTIGUOUS(x)
+
+#define CHECK_DIM(d, x) TORCH_CHECK(x.dim() == d, #x " must be a " #d "D tensor")
+
+#define CHECK_EQ(a, b) TORCH_CHECK((a) == (b), "CHECK_EQ(" #a ", " #b ") failed. ", a, " vs ", b)
+
+// parallel routines
+constexpr int GRAIN_SIZE = 1024;
+
+template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
+inline T div_up(T x, T y) { return (x + y - 1) / y; }
+
+template <typename T>
+inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
+#if 0
+    // onednn partition pattern
+    T& n_my = n_end;
+    if (nth <= 1 || n == 0) {
+        n_start = 0;
+        n_my = n;
+    } else {
+        T n1 = div_up(n, nth);
+        T n2 = n1 - 1;
+        T T1 = n - n2 * nth;
+        n_my = ith < T1 ? n1 : n2;
+        n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2;
+    }
+    n_end += n_start;
+#else
+    // pytorch aten partition pattern
+    T n_my = div_up(n, nth);
+    n_start = ith * n_my;
+    n_end = std::min(n_start + n_my, n);
+#endif
+}
+
+template <typename func_t>
+inline void parallel_for(int n, const func_t& f) {
+#if defined(_OPENMP)
+#pragma omp parallel
+{
+    int nth = omp_get_num_threads();
+    int ith = omp_get_thread_num();
+    int tbegin, tend;
+    balance211(n, nth, ith, tbegin, tend);
+    f(tbegin, tend);
+}
+#else
+    f(0, n);
+#endif
+}
+
+// for 1d parallel, use `actual_nth`
+// for 2d parallel, use even nths, e.g. 43->42
+int inline adjust_num_threads(int m) {
+  int actual_nth = at::get_num_threads();
+  if (m == 1) {
+    return actual_nth;
+  }
+  return std::max(1, (actual_nth >> 1) * 2);
+}
+
+template <typename func_t>
+inline void parallel_2d(int m, int n, const func_t& f) {
+
+  // make sure we have even num_threads
+  int nth = adjust_num_threads(m);
+
+  // [NOTE] thread blocking:
+  //
+  //   1) prefer square block per thread
+  //   2) use even number of CPU cores
+  //   3) use all `num_threads` cores
+  //
+  //   we have:
+  //     TM * TN = T
+  //     BM / TM = BN / TN
+  //   then:
+  //     TM = ((BM / BN) * T) ^ 0.5
+  //
+  float r = float(m) / n;
+  int nth_m = std::ceil(std::sqrt(r * nth));
+  int nth_n = 1;
+  for (; nth_m > 0; --nth_m) {
+    nth_n = nth / nth_m;
+    if (nth_m * nth_n == nth) {
+      break;
+    }
+  }
+
+#if defined(_OPENMP)
+#pragma omp parallel num_threads(nth)
+{
+  int ith = omp_get_thread_num();
+  int ith_m = ith / nth_n;
+  int ith_n = ith % nth_n;
+
+  int thread_block_m = div_up(m, nth_m);
+  int thread_block_n = div_up(n, nth_n);
+
+  int begin_m = ith_m * thread_block_m;
+  int end_m = std::min(m, begin_m + thread_block_m);
+  int begin_n = ith_n * thread_block_n;
+  int end_n = std::min(n, begin_n + thread_block_n);
+
+  f(begin_m, end_m, begin_n, end_n);
+}
+#else
+  f(0, m, 0, n);
+#endif
+}
+
+template <typename T>
+int get_cache_blocks(int BLOCK_SIZE, int K) {
+  // L2 2MB and ratio of 50%
+  const int L2_size = 2048 * 1024 >> 1;
+  return std::max(1, int(L2_size / (BLOCK_SIZE * K * sizeof(T))));
+}
+
+// data indexing for dimension collapse
+template <typename T>
+inline T data_index_init(T offset) {
+  return offset;
+}
+
+template <typename T, typename... Args>
+inline T data_index_init(T offset, T& x, const T& X, Args&&... args) {
+  offset = data_index_init(offset, std::forward<Args>(args)...);
+  x = offset % X;
+  return offset / X;
+}
+
+inline bool data_index_step() {
+  return true;
+}
+
+template <typename T, typename... Args>
+inline bool data_index_step(T& x, const T& X, Args&&... args) {
+  if (data_index_step(std::forward<Args>(args)...)) {
+    x = ((x + 1) == X) ? 0 : (x + 1);
+    return x == 0;
+  }
+  return false;
+}
+
+// forced unroll for perf critical path
+
+#if __has_attribute(always_inline)
+#define ALWAYS_INLINE __attribute__((__always_inline__)) inline
+#else
+#define ALWAYS_INLINE inline
+#endif
+
+template <int n>
+struct Unroll {
+  template <typename Func, typename... Args>
+  ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
+    Unroll<n - 1>{}(f, args...);
+    f(std::integral_constant<int, n - 1>{}, args...);
+  }
+};
+
+template <>
+struct Unroll<1> {
+  template <typename Func, typename... Args>
+  ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
+    f(std::integral_constant<int, 0>{}, args...);
+  }
+};
+
+} // anonymous namespace

csrc/cpu/sgl-kernels/gemm.cpp

@@ -0,0 +1,464 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+namespace {
+
+// packed   layout:
+//   quants {N, K}  int8_t
+//   comp   {N}     int32_t
+template <int BLOCK_N>
+inline void s8s8_compensation(int8_t* __restrict__ packed, int K) {
+#if defined(CPU_CAPABILITY_AVX512)
+  constexpr int COLS = BLOCK_N / 16;
+  __m512i vcomp[COLS];
+
+  for (int col = 0; col < COLS; ++col) {
+    vcomp[col] = _mm512_setzero_si512();
+  }
+
+  const int64_t offset = BLOCK_N * K;
+  const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
+  for (int k = 0; k < K / 4; ++k) {
+    for (int col = 0; col < COLS; ++col) {
+      __m512i vb = _mm512_loadu_si512((const __m512i *)(packed + k * BLOCK_N * 4 + col * 64));
+      vcomp[col] = _mm512_dpbusd_epi32(vcomp[col], off, vb);
+    }
+  }
+
+  for (int col = 0; col < COLS; ++col) {
+    _mm512_storeu_si512((__m512i *)(packed + offset + col * 64), vcomp[col]);
+  }
+#else
+  TORCH_CHECK(false, "s8s8_compensation not implemented!");
+#endif
+}
+
+// convert to vnni format
+// from [N, K] to [K/2, N, 2] for bfloat16 and float16
+template <typename packed_t>
+inline void pack_vnni(packed_t* __restrict__ packed, const packed_t* __restrict__ weight, int N, int K) {
+  const int VNNI_BLK = 2;
+  for (int n = 0; n < N; ++n) {
+    for (int k = 0; k < K / VNNI_BLK; ++k) {
+      for (int d = 0; d < VNNI_BLK; ++d) {
+        packed[k * N * VNNI_BLK + n * VNNI_BLK + d] = weight[n * K + k * VNNI_BLK + d];
+      }
+    }
+  }
+}
+
+template <>
+inline void pack_vnni<int8_t>(int8_t* __restrict__ packed, const int8_t* __restrict__ weight, int N, int K) {
+  constexpr int BLOCK_N = block_size_n();
+  TORCH_CHECK(N == BLOCK_N);
+
+  const int VNNI_BLK = 4;
+  for (int n = 0; n < N; ++n) {
+    for (int k = 0; k < K / VNNI_BLK; ++k) {
+      for (int d = 0; d < VNNI_BLK; ++d) {
+        packed[k * N * VNNI_BLK + n * VNNI_BLK + d] = weight[n * K + k * VNNI_BLK + d];
+      }
+    }
+  }
+  s8s8_compensation<BLOCK_N>(packed, K);
+}
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d]);
+  }
+}
+
+template <typename scalar_t>
+inline void copy_add_stub(scalar_t* __restrict__ out, const float* __restrict__ input, const float* __restrict__ bias, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d) + fVec::loadu(bias + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size()) + fVec::loadu(bias + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + bias[d]);
+  }
+}
+
+template <typename scalar_t, bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn {
+  static inline void apply(
+      const scalar_t* __restrict__ A, const scalar_t* __restrict__ B, scalar_t* __restrict__ C,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn<at::BFloat16, has_bias, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const at::BFloat16* __restrict__ A, const at::BFloat16* __restrict__ B, at::BFloat16* __restrict__ C,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+
+    // prefetch distance
+    constexpr int PREFETCH_SIZE_K = 0;
+
+    __m512bh va;
+    __m512bh vb[COLS];
+    __m512 vc[ROWS * COLS];
+
+    auto loadc = [&](auto i) {
+      constexpr int col = i % COLS;
+      if constexpr (has_bias) {
+        vc[i] = _mm512_loadu_ps(bias + col * 16);
+      } else {
+        vc[i] = _mm512_set1_ps(0.f);
+      }
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K2 = K >> 1;
+    const int64_t lda2 = lda >> 1;
+    const int64_t ldb2 = ldb; // ldb * 2 >> 1;
+    const float* a_ptr = reinterpret_cast<const float*>(A);
+    const float* b_ptr = reinterpret_cast<const float*>(B);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = (__m512bh)(_mm512_set1_ps(a_ptr[row * lda2 + k]));
+      }
+      if constexpr (row == 0) {
+        vb[col] = (__m512bh)(_mm512_loadu_si512(b_ptr + k * ldb2 + col * 16));
+        if constexpr (PREFETCH_SIZE_K > 0) {
+          _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2 + col * 16, _MM_HINT_T0);
+        }
+      }
+      vc[i] = _mm512_dpbf16_ps(vc[i], va, vb[col]);
+    };
+    for (int64_t k = 0; k < K2; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+      // for COLS = 2, 4 use 512bit store
+      // for COLS = 1, 3 use 256bit store
+      if constexpr (COLS % 2 == 0) {
+        if constexpr (col % 2 == 0) {
+          _mm512_storeu_si512(
+              reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+              (__m512i)(_mm512_cvtne2ps_pbh(vc[row * COLS + col + 1], vc[row * COLS + col])));
+        }
+      } else {
+        _mm256_storeu_si256(
+            reinterpret_cast<__m256i*>(C + row * ldc + col * 16),
+            (__m256i)(_mm512_cvtneps_pbh(vc[i])));
+      }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE)                          \
+    tinygemm_kernel_nn<scalar_t, has_bias, MB_SIZE, NB_SIZE>::apply(         \
+        A + mb_start * lda, B + nb_start * 2, C + mb_start * ldc + nb_start, \
+        has_bias ? bias + nb_start : nullptr, K, lda, ldb, ldc);
+
+template <typename scalar_t, bool has_bias>
+struct brgemm {
+  static inline void apply(
+      const scalar_t* __restrict__ A, const scalar_t* __restrict__ B, scalar_t* __restrict__ C,
+      float* __restrict__ Ctmp, const float* __restrict__ bias,
+      int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int BLOCK_N = block_size_n();
+    at::native::cpublas::brgemm(
+        M, N, K, lda, ldb, BLOCK_N, /* add_C */false,
+        A, B, Ctmp);
+
+    // copy from Ctmp to C
+    for (int64_t m = 0; m < M; ++m) {
+      if constexpr (has_bias) {
+        copy_add_stub(C + m * ldc, Ctmp + m * BLOCK_N, bias, N);
+      } else {
+        copy_stub(C + m * ldc, Ctmp + m * BLOCK_N, N);
+      }
+    }
+  }
+};
+
+template <typename scalar_t, bool has_bias>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const scalar_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg) {
+
+  if (brg) {
+    brgemm<scalar_t, has_bias>::apply(
+        A, B, C, Ctmp, bias,
+        M, N, K, lda, ldb, ldc);
+    return;
+  }
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        // mb_size = 1
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_NN(1, 32); break;
+        case 0x14: LAUNCH_TINYGEMM_KERNEL_NN(1, 64); break;
+        // mb_size = 2
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_NN(2, 32); break;
+        case 0x24: LAUNCH_TINYGEMM_KERNEL_NN(2, 64); break;
+        // mb_size = 3
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_NN(3, 32); break;
+        case 0x34: LAUNCH_TINYGEMM_KERNEL_NN(3, 64); break;
+        // mb_size = 4
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_NN(4, 32); break;
+        case 0x44: LAUNCH_TINYGEMM_KERNEL_NN(4, 64); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+void weight_packed_linear_kernel_impl(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ mat1,
+    const scalar_t* __restrict__ mat2,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t mat1_strideM,
+    int64_t out_strideM) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  // use avx512-bf16 when a) M is small; b) dtype is bfloat16, otherwise use amx
+  const bool use_brgemm = (M > 4) || (!std::is_same_v<scalar_t, at::BFloat16>);
+
+  // l2 cache block for n
+  int64_t cache_blocks_nb = get_cache_blocks<scalar_t>(BLOCK_N, K);
+
+  // parallel on [MB, NB]
+  AT_DISPATCH_BOOL(bias != nullptr, has_bias, [&] {
+    parallel_2d(MB, NB, [&](int64_t begin_mb, int64_t end_mb, int64_t begin_nb, int64_t end_nb) {
+
+      // for brgemm, use float32 for accumulate
+      alignas(64) float Ctmp[BLOCK_M * BLOCK_N];
+
+      for (int64_t nbb = begin_nb; nbb < end_nb; nbb += cache_blocks_nb) {
+      for (int64_t mb = begin_mb; mb < end_mb; ++mb) {
+      for (int64_t nb = nbb; nb < std::min(nbb + cache_blocks_nb, end_nb); ++nb) {
+
+        int64_t mb_start = mb * BLOCK_M;
+        int64_t mb_size = std::min(M - mb_start, BLOCK_M);
+        int64_t nb_start = nb * BLOCK_N;
+        int64_t nb_size = std::min(N - nb_start, BLOCK_N);
+
+        tinygemm_kernel<scalar_t, has_bias>(
+            /*   A */ mat1 + mb_start * mat1_strideM,
+            /*   B */ mat2 + nb_start * K /* nb * BLOCK_N * K */,
+            /*   C */ out + mb_start * out_strideM + nb_start,
+            /* Ctmp*/ Ctmp,
+            /* bias*/ bias + nb_start,
+            /*   M */ mb_size,
+            /*   N */ nb_size,
+            /*   K */ K,
+            /* lda */ mat1_strideM,
+            /* ldb */ nb_size,
+            /* ldc */ out_strideM,
+            /* brg */ use_brgemm);
+      }}}
+
+      if (use_brgemm) {
+        at::native::cpublas::brgemm_release();
+      }
+    });
+  });
+}
+
+} // anonymous namespace
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(const scalar_t* __restrict__ A, const scalar_t* __restrict__ B, scalar_t* __restrict__ C,
+    float* __restrict__ Ctmp, int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc, bool brg) {
+  tinygemm_kernel<scalar_t, false>(A, B, C, Ctmp, nullptr, M, N, K, lda, ldb, ldc, brg);
+}
+
+#define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE)                                             \
+    template void tinygemm_kernel<TYPE>(                                                \
+        const TYPE* __restrict__ A, const TYPE* __restrict__ B, TYPE* __restrict__ C,   \
+        float* __restrict__ Ctmp, int64_t M, int64_t N, int64_t K, int64_t lda,         \
+        int64_t ldb, int64_t ldc, bool brg)
+
+INSTANTIATE_TINYGEMM_TEMPLATE(at::BFloat16);
+INSTANTIATE_TINYGEMM_TEMPLATE(at::Half);
+
+at::Tensor convert_weight_packed(at::Tensor& weight) {
+  // for 3d moe weights
+  // weight : [E, OC, IC]
+  //     w1 : [E, 2N,  K]
+  //     w2 : [E,  K,  N]
+  CHECK_INPUT(weight);
+
+  const int64_t ndim = weight.ndimension();
+  TORCH_CHECK(ndim == 2 || ndim == 3, "expect weight to be 2d or 3d, got ", ndim, "d tensor.");
+  const auto st = weight.scalar_type();
+  const int64_t E = ndim == 3 ? weight.size(0) : 1;
+  const int64_t OC = ndim == 3 ? weight.size(1) : weight.size(0);
+  const int64_t IC = ndim == 3 ? weight.size(2) : weight.size(1);
+
+  // we handle 2 TILE_N at a time.
+  TORCH_CHECK(OC % TILE_N == 0, "invalid weight out features ", OC);
+  TORCH_CHECK(IC % TILE_K == 0, "invalid weight input features ", IC);
+
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t NB = div_up(OC, BLOCK_N);
+
+  // use phony sizes here [E, OC, IC], for each [E], [OC, IC] -> [IC / 2, OC, 2]
+  auto packed_weight = at::empty({}, weight.options());
+  const int64_t stride = OC * IC;
+
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf || st == at::kChar || st == at::kFloat8_e4m3fn,
+      "expect weight to be bfloat16, float16, int8 or fp8_e4m3.");
+
+  CPU_DISPATCH_PACKED_TYPES(st, [&] {
+    // adjust most inner dimension size
+    const int packed_row_size = get_row_size<packed_t>(IC);
+    auto sizes = weight.sizes().vec();
+    sizes[ndim - 1] = packed_row_size;
+    packed_weight.resize_(sizes);
+
+    const packed_t* w_data = weight.data_ptr<packed_t>();
+    packed_t* packed_data = packed_weight.data_ptr<packed_t>();
+
+    // parallel on {E, NB}
+    at::parallel_for(0, E * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t e{0}, nb{0};
+      data_index_init(begin, e, E, nb, NB);
+
+      for (int64_t i = begin; i < end; ++i) {
+        UNUSED(i);
+
+        int64_t n = nb * BLOCK_N;
+        int64_t n_size = std::min(BLOCK_N, OC - n);
+        pack_vnni<packed_t>(
+            packed_data + e * OC * packed_row_size + n * packed_row_size,
+            w_data + e * stride + n * IC,
+            n_size,
+            IC);
+
+        // move to the next index
+        data_index_step(e, E, nb, NB);
+      }
+    });
+  });
+  return packed_weight;
+}
+
+// mat1 : [M, K]
+// mat2 : [N, K]
+// bias : [N]
+// out  : [M, N]
+//
+at::Tensor weight_packed_linear(at::Tensor& mat1, at::Tensor& mat2,
+    const std::optional<at::Tensor>& bias, bool is_vnni) {
+  RECORD_FUNCTION(
+    "sgl-kernel::weight_packed_linear", std::vector<c10::IValue>({mat1, mat2, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(mat1);
+  CHECK_INPUT(mat2);
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat2.size(1);
+  CHECK_EQ(mat1.size(1), K);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  auto out = at::empty({M, N}, mat1.options());
+
+  // strides
+  int64_t mat1_strideM = mat1.stride(0);
+  int64_t out_strideM = out.stride(0);
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(mat1.scalar_type(), "weight_packed_linear_kernel_impl", [&] {
+    weight_packed_linear_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        mat1.data_ptr<scalar_t>(),
+        packed_w.data_ptr<scalar_t>(),
+        bias_data,
+        M,
+        N,
+        K,
+        mat1_strideM,
+        out_strideM);
+  });
+
+  return out;
+}

csrc/cpu/sgl-kernels/gemm.h

@@ -0,0 +1,266 @@
+#pragma once
+
+#include <ATen/native/CPUBlas.h>
+
+// clang-format off
+
+// amx-bf16
+#define TILE_M 16
+#define TILE_N 16
+#define TILE_K 32
+
+// block size for AMX gemm
+constexpr int block_size_m() { return 2 * TILE_M; }
+constexpr int block_size_n() { return 2 * TILE_N; }
+
+// define threshold using brgemm (intel AMX)
+template <typename T> inline bool can_use_brgemm(int M);
+template <> inline bool can_use_brgemm<at::BFloat16>(int M) { return M > 4; }
+template <> inline bool can_use_brgemm<at::Half>(int M) { return true; }
+// TODO: add u8s8 brgemm, this requires PyTorch 2.7
+template <> inline bool can_use_brgemm<int8_t>(int M) { return false; }
+template <> inline bool can_use_brgemm<at::Float8_e4m3fn>(int M) { return M > 4; }
+template <> inline bool can_use_brgemm<at::quint4x2>(int M) { return M > 4; }
+
+// work around compiler internal error
+#define BLOCK_K 128 // 4 * TILE_K
+
+// adjust leading dimension size for K
+template <typename T>
+inline int64_t get_row_size(int64_t K) {
+  return K;
+}
+
+template <>
+inline int64_t get_row_size<int8_t>(int64_t K) {
+  return K + sizeof(int32_t);
+}
+
+inline int64_t get_row_size(int64_t K, bool use_int8_w8a8) {
+  return use_int8_w8a8 ? K + sizeof(int32_t) : K;
+}
+
+// pack weight to vnni format
+at::Tensor convert_weight_packed(at::Tensor& weight);
+
+// moe implementations for int8 w8a8
+template <typename scalar_t>
+void fused_experts_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    uint8_t* __restrict__ A_tmp,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad);
+
+// moe implementations for fp8 w8a16
+template <typename scalar_t>
+void fused_experts_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    scalar_t* __restrict__ A_tmp,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad);
+
+// moe implementations for int4 w4a16
+template <typename scalar_t>
+void fused_experts_int4_w4a16_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    scalar_t* __restrict__ A_tmp,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::quint4x2* __restrict__ packed_w1,
+    const at::quint4x2* __restrict__ packed_w2,
+    const uint8_t* __restrict__ w1z,
+    const uint8_t* __restrict__ w2z,
+    const scalar_t* __restrict__ w1s,
+    const scalar_t* __restrict__ w2s,
+    int group_size,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad);
+
+// shared expert implememntation for int8 w8a8
+template <typename scalar_t>
+void shared_expert_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K);
+
+template <typename scalar_t>
+void shared_expert_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K);
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const scalar_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    float* __restrict__ Ctmp,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    int32_t* __restrict__ Ctmp,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::Float8_e4m3fn* __restrict__ B,
+    scalar_t* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ scale,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::quint4x2* __restrict__ B,
+    scalar_t* __restrict__ C,
+    const uint8_t* __restrict__ Bz,
+    const scalar_t* __restrict__ Bs,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int group_size,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    int64_t strideBz,
+    int64_t strideBs,
+    bool brg);
+
+// TODO: debug print, remove me later
+inline void print_16x32i(const __m512i x) {
+  int32_t a[16];
+  _mm512_storeu_si512((__m512i *)a, x);
+
+  for (int i = 0; i < 16; i++){
+    std::cout << a[i] << " ";
+  }
+  std::cout << std::endl;
+}
+
+inline void print_16x32(const __m512 x) {
+  float a[16];
+  _mm512_storeu_ps((__m512 *)a, x);
+
+  for (int i = 0; i < 16; i++){
+    std::cout << a[i] << " ";
+  }
+  std::cout << std::endl;
+}
+
+
+inline void print_32x8u(const __m256i x) {
+  uint8_t a[32];
+  _mm256_storeu_si256((__m256i *)a, x);
+
+  for (int i = 0; i < 32; ++i) {
+    std::cout << int32_t(a[i]) << " ";
+  }
+  std::cout << std::endl;
+}

csrc/cpu/sgl-kernels/gemm_fp8.cpp

@@ -0,0 +1,530 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+// we use 4x32 for BLOCK_M
+#define BLOCK_SIZE_M_SCALE 4
+
+namespace {
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d]);
+  }
+}
+
+template <typename scalar_t>
+inline void copy_add_stub(scalar_t* __restrict__ out, const float* __restrict__ input, const float* __restrict__ bias, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d) + fVec::loadu(bias + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size()) + fVec::loadu(bias + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + bias[d]);
+  }
+}
+
+inline void unpack_B(
+    at::BFloat16* __restrict__ Btmp,
+    const at::Float8_e4m3fn* __restrict__ packed_B,
+    int N,
+    int K,
+    int ldb,
+    int ldb_tmp,
+    float scale) {
+#if defined(CPU_CAPABILITY_AVX512)
+  // [K/2, N, 2]
+  const int K2 = K >> 1;
+  const int ldb2 = ldb; // ldb * 2 >> 1;
+  const uint16_t* b_ptr = reinterpret_cast<const uint16_t*>(packed_B);
+  const __m512 vd = _mm512_set1_ps(scale);
+
+  constexpr int BLOCK_N = block_size_n();
+  static_assert(BLOCK_N == 32);
+
+  // prefetch distance
+  constexpr int PREFETCH_SIZE_K = 64;
+
+#pragma GCC unroll 4
+  for (int k = 0; k < K2; ++k) {
+    __m512i b8 = _mm512_loadu_si512(b_ptr + k * ldb2);
+    if constexpr (PREFETCH_SIZE_K > 0) {
+      _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2, _MM_HINT_T0);
+    }
+
+    __m256i b8_0 = _mm512_extracti32x8_epi32(b8, 0);
+    __m256i b8_1 = _mm512_extracti32x8_epi32(b8, 1);
+
+    __m512bh bf16_0 = CVT_FP8_TO_BF16(b8_0);
+    __m512bh bf16_1 = CVT_FP8_TO_BF16(b8_1);
+
+    // Apply scale
+    __m512 f0_lo = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_0, 0));
+    __m512 f0_hi = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_0, 1));
+    __m512 f1_lo = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_1, 0));
+    __m512 f1_hi = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_1, 1));
+
+    f0_lo = _mm512_mul_ps(f0_lo, vd);
+    f0_hi = _mm512_mul_ps(f0_hi, vd);
+    f1_lo = _mm512_mul_ps(f1_lo, vd);
+    f1_hi = _mm512_mul_ps(f1_hi, vd);
+
+    bf16_0 = _mm512_cvtne2ps_pbh(f0_hi, f0_lo);
+    bf16_1 = _mm512_cvtne2ps_pbh(f1_hi, f1_lo);
+
+    _mm512_storeu_si512(Btmp + k * ldb_tmp * 2 + 0, (__m512i)bf16_0);
+    _mm512_storeu_si512(Btmp + k * ldb_tmp * 2 + 32, (__m512i)bf16_1);
+  }
+#else
+  TORCH_CHECK(false, "unpack_B: scalar path not implemented!");
+#endif
+}
+
+template <typename scalar_t, typename packed_t, bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn {
+  static inline void apply(
+      const scalar_t* __restrict__ A, const packed_t* __restrict__ B, scalar_t* __restrict__ C,
+      const float* __restrict__ bias, const float* __restrict__ scale, int K, int lda, int ldb, int ldc, int64_t block_size_K) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn<at::BFloat16, at::Float8_e4m3fn, has_bias, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const at::BFloat16* __restrict__ A, const at::Float8_e4m3fn* __restrict__ B, at::BFloat16* __restrict__ C,
+      const float* __restrict__ bias, const float* __restrict__ scale, int K, int lda, int ldb, int ldc, int64_t block_size_K) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+
+    const int KB = div_up(K, BLOCK_K);
+
+    // prefetch distance
+    constexpr int PREFETCH_SIZE_K = 64;
+    constexpr int PREFETCH_SIZE_KB = 1;
+
+    __m512bh va;
+    __m512bh vb[COLS];
+    __m512 vc[ROWS * COLS];
+    __m512 vsum[ROWS * COLS];
+
+    // block quant scale
+    __m512 vscale;
+
+    auto loadc = [&](auto i) {
+      constexpr int col = i % COLS;
+      if constexpr (has_bias) {
+        vc[i] = _mm512_loadu_ps(bias + col * 16);
+      } else {
+        vc[i] = _mm512_setzero_ps();
+      }
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int lda2 = lda >> 1;
+    const int ldb2 = ldb; // ldb * 2 >> 1;
+    const float* a_ptr = reinterpret_cast<const float*>(A);
+    const uint16_t* b_ptr = reinterpret_cast<const uint16_t*>(B);
+
+    auto compute = [&](auto i, int k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = (__m512bh)(_mm512_set1_ps(a_ptr[row * lda2 + k]));
+        if constexpr (PREFETCH_SIZE_K > 0) {
+          _mm_prefetch(a_ptr + row * lda2 + k + PREFETCH_SIZE_K, _MM_HINT_T0);
+        }
+      }
+      if constexpr (row == 0) {
+        if constexpr (col % 2 == 0) {
+          __m512i b8 = _mm512_loadu_si512(b_ptr + k * ldb2 + col * 16);
+          if constexpr (PREFETCH_SIZE_K > 0) {
+            _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2 + col * 16, _MM_HINT_T0);
+          }
+          vb[col + 0] = CVT_FP8_TO_BF16(_mm512_extracti32x8_epi32(b8, 0));
+          vb[col + 1] = CVT_FP8_TO_BF16(_mm512_extracti32x8_epi32(b8, 1));
+        }
+      }
+      vsum[i] = _mm512_dpbf16_ps(vsum[i], va, vb[col]);
+    };
+
+    constexpr int BLOCK_K2 = BLOCK_K >> 1;
+    for (int kb = 0; kb < KB; ++kb) {
+      int kb_start = kb * BLOCK_K2;
+      int kb_end = std::min(K, kb_start + BLOCK_K2);
+      // 1. load scale vector
+      vscale = _mm512_set1_ps(scale[kb]);
+      if constexpr (PREFETCH_SIZE_KB > 0) {
+        _mm_prefetch(scale + kb + PREFETCH_SIZE_KB, _MM_HINT_T0);
+      }
+      // 2. zero vsum for each block
+      Unroll<ROWS * COLS>{}([&](auto i) {
+        vsum[i] = _mm512_setzero_ps();
+      });
+      // 3. accumulate across each block
+      for (int k = kb_start; k < kb_end; ++k) {
+        Unroll<ROWS * COLS>{}(compute, k);
+      }
+      // 4. apply scale
+      Unroll<ROWS * COLS>{}([&](auto i) {
+        vc[i] = _mm512_fmadd_ps(vsum[i], vscale, vc[i]);
+      });
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+      // for COLS = 2,4 use 512bit store
+      if constexpr (col % 2 == 0) {
+        _mm512_storeu_si512(
+            reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+            (__m512i)(_mm512_cvtne2ps_pbh(vc[row * COLS + col + 1], vc[row * COLS + col])));
+      }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE)                          \
+    tinygemm_kernel_nn<scalar_t, at::Float8_e4m3fn, has_bias, MB_SIZE, NB_SIZE>::apply(         \
+        A + mb_start * lda, B + nb_start * 2, C + mb_start * ldc + nb_start, \
+        has_bias ? bias + nb_start : nullptr, scale, K, lda, ldb, ldc, block_size_K);
+
+template <typename scalar_t, typename packed_t, bool has_bias>
+struct brgemm {
+  static inline void apply(
+      const scalar_t* __restrict__ A,
+      const packed_t* __restrict__ B,
+      scalar_t* __restrict__ C,
+      scalar_t* __restrict__ Btmp,
+      float* __restrict__ Ctmp,
+      const float* __restrict__ bias,
+      const float* __restrict__ scale,
+      int M,
+      int N,
+      int K,
+      int lda,
+      int ldb,
+      int ldc) {
+    TORCH_CHECK(false, "struct brgemm: primary template not implemented!");
+  }
+};
+
+template <bool has_bias>
+struct brgemm<at::BFloat16, at::Float8_e4m3fn, has_bias> {
+  static inline void apply(
+      const at::BFloat16* __restrict__ A,
+      const at::Float8_e4m3fn* __restrict__ B,
+      at::BFloat16* __restrict__ C,
+      at::BFloat16* __restrict__ Btmp,
+      float* __restrict__ Ctmp,
+      const float* __restrict__ bias,
+      const float* __restrict__ scale,
+      int M,
+      int N,
+      int K,
+      int lda,
+      int ldb,
+      int ldc) {
+
+    constexpr int BLOCK_N = block_size_n();
+
+    // [K, BLOCK_N] -> [K / 2, BLOCK_N * 2]
+    const int ldb_tmp = BLOCK_N;
+
+    for (int k = 0; k < K; k += BLOCK_K) {
+      int kb_size = std::min(BLOCK_K, K - k);
+
+      int idx = k >> 7; // k / BLOCK_K where BLOCK_K = 128
+      unpack_B(Btmp + k * ldb_tmp, B + k * ldb, N, kb_size, ldb, ldb_tmp, scale[idx]);
+    }
+
+    at::native::cpublas::brgemm(
+        M, N, K, lda, ldb_tmp, BLOCK_N, /* add_C */ false, A, Btmp, Ctmp);
+
+    // copy from Ctmp to C
+    for (int m = 0; m < M; ++m) {
+      if constexpr (has_bias) {
+        copy_add_stub(C + m * ldc, Ctmp + m * BLOCK_N, bias, N);
+      } else {
+        copy_stub(C + m * ldc, Ctmp + m * BLOCK_N, N);
+      }
+    }
+  }
+};
+
+template <typename scalar_t, bool has_bias>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::Float8_e4m3fn* __restrict__ B,
+    scalar_t* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ scale,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K) {
+
+  if (brg) {
+    brgemm<scalar_t, at::Float8_e4m3fn, has_bias>::apply(
+        A, B, C, Btmp, Ctmp, bias, scale, M, N, K, lda, ldb, ldc);
+    return;
+  }
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_NN(1, 32); break;
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_NN(2, 32); break;
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_NN(3, 32); break;
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_NN(4, 32); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+void fp8_scaled_mm_kernel_impl(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ mat1,
+    const at::Float8_e4m3fn* __restrict__ mat2,
+    const float* __restrict__ scales2,
+    const float* __restrict__ bias,
+    scalar_t* __restrict__ buffer,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t mat1_strideM,
+    int64_t out_strideM,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    int64_t buffer_size_per_thread) {
+
+  constexpr int64_t BLOCK_M = block_size_m() * BLOCK_SIZE_M_SCALE;
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  const int64_t scale_size_K = div_up(K, block_size_K);
+  const int64_t blocks_n_per_group = block_size_N / BLOCK_N;
+
+  const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(M);
+
+  // parallel on [MB, NB]
+  AT_DISPATCH_BOOL(bias != nullptr, has_bias, [&] {
+    at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t mb{0}, nb{0};
+      data_index_init(begin, mb, MB, nb, NB);
+
+      int tid = at::get_thread_num();
+      scalar_t* __restrict__ Btmp = buffer + tid * buffer_size_per_thread;
+      float* __restrict__ Ctmp = (float*)((void*)(Btmp + BLOCK_N * K));
+
+      for (int64_t i = begin; i < end; ++i) {
+        UNUSED(i);
+        const float* scale_ptr = scales2 + (nb / blocks_n_per_group) * scale_size_K;
+
+        int64_t mb_start = mb * BLOCK_M;
+        int64_t mb_size = std::min(M - mb_start, BLOCK_M);
+        int64_t nb_start = nb * BLOCK_N;
+        int64_t nb_size = std::min(N - nb_start, BLOCK_N);
+
+        tinygemm_kernel<scalar_t, has_bias>(
+            /*   A            */ mat1 + mb_start * mat1_strideM,
+            /*   B            */ mat2 + nb_start * K, // nb * BLOCK_N * K
+            /*   C            */ out + mb_start * out_strideM + nb_start,
+            /*   Btmp         */ Btmp,
+            /*   Ctmp         */ Ctmp,
+            /*   scale        */ scale_ptr,
+            /*   bias         */ bias + nb_start,
+            /*   M            */ mb_size,
+            /*   N            */ nb_size,
+            /*   K            */ K,
+            /*   lda          */ mat1_strideM,
+            /*   ldb          */ nb_size,
+            /*   ldc          */ out_strideM,
+            /*   brg          */ use_brgemm,
+            /*   block_size_K */ block_size_K);
+
+        // move to the next index
+        data_index_step(mb, MB, nb, NB);
+      }
+
+      if (use_brgemm) {
+        at::native::cpublas::brgemm_release();
+      }
+    });
+  });
+}
+
+} // anonymous namespace
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::Float8_e4m3fn* __restrict__ B,
+    scalar_t* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ scale,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K) {
+  tinygemm_kernel<scalar_t, false>(A, B, C, Btmp, Ctmp, scale, nullptr, M, N, K, lda, ldb, ldc, brg, block_size_K);
+}
+
+#define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE)    \
+  template void tinygemm_kernel<TYPE>(         \
+      const TYPE* __restrict__ A,              \
+      const at::Float8_e4m3fn* __restrict__ B, \
+      TYPE* __restrict__ C,                    \
+      TYPE* __restrict__ Btmp,                 \
+      float* __restrict__ Ctmp,                \
+      const float* __restrict__ scale,         \
+      int64_t M,                               \
+      int64_t N,                               \
+      int64_t K,                               \
+      int64_t lda,                             \
+      int64_t ldb,                             \
+      int64_t ldc,                             \
+      bool brg,                                \
+      int64_t block_size_K)
+
+INSTANTIATE_TINYGEMM_TEMPLATE(at::BFloat16);
+INSTANTIATE_TINYGEMM_TEMPLATE(at::Half);
+
+at::Tensor fp8_scaled_mm_cpu(at::Tensor& mat1, at::Tensor& mat2, at::Tensor& scales2,
+    std::vector<int64_t> block_size, std::optional<at::Tensor>& bias,
+    at::ScalarType out_dtype, bool is_vnni) {
+  RECORD_FUNCTION("sgl-kernel::fp8_scaled_mm_cpu", std::vector<c10::IValue>({mat1, mat2, scales2, block_size, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(mat1);
+  CHECK_INPUT(mat2);
+  CHECK_INPUT(scales2);
+  TORCH_CHECK(scales2.scalar_type() == at::kFloat,
+      "fp8_scaled_mm_cpu: expect scales2 to be float32.");
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat2.size(1);
+
+  CHECK_EQ(mat1.size(1), K);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  TORCH_CHECK(block_size.size() == 2,
+      "fp8_scaled_mm_cpu: expect block_size.size() to be 2.");
+
+  int64_t block_size_N = block_size[0];
+  int64_t block_size_K = block_size[1];
+
+  constexpr int64_t BLOCK_M = block_size_m() * BLOCK_SIZE_M_SCALE;
+  constexpr int64_t BLOCK_N = block_size_n();
+  TORCH_CHECK(block_size_N % BLOCK_N == 0, "fp8_scaled_mm_cpu: expect block_size_N to be multiples of BLOCK_N");
+  TORCH_CHECK(block_size_K == BLOCK_K, "fp8_scaled_mm_cpu: expect block_size_K equals to BLOCK_K");
+  CHECK_EQ(scales2.size(0), div_up(N, block_size_N));
+  CHECK_EQ(scales2.size(1), div_up(K, block_size_K));
+
+  const auto st = mat1.scalar_type();
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf,
+      "fp8_scaled_mm_cpu: expect A to be bfloat16 or half.");
+  TORCH_CHECK(st == out_dtype,
+      "fp8_scaled_mm_cpu: expect A has same dtype with out_dtype.");
+  TORCH_CHECK(mat2.scalar_type() == at::kFloat8_e4m3fn,
+      "fp8_scaled_mm_cpu: expect mat2 to be fp8_e4m3.");
+  TORCH_CHECK(scales2.scalar_type() == at::kFloat,
+      "fp8_scaled_mm_cpu: expect scales to be float32.");
+  auto out = at::empty({M, N}, mat1.options().dtype(out_dtype));
+
+  // strides
+  int64_t mat1_strideM = mat1.stride(0);
+  int64_t out_strideM = out.stride(0);
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  // Btmp : [T, BLOCK_N * K]
+  // Ctmp : [T, BLOCK_M * BLOCK_N]
+  int num_threads = at::get_num_threads();
+  int64_t size_per_thread = BLOCK_N * K + BLOCK_M * BLOCK_N * 2;
+  auto buffer = at::empty({num_threads, size_per_thread}, mat1.options());
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "fp8_scaled_mm_kernel_impl", [&] {
+    fp8_scaled_mm_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        mat1.data_ptr<scalar_t>(),
+        packed_w.data_ptr<at::Float8_e4m3fn>(),
+        scales2.data_ptr<float>(),
+        bias_data,
+        buffer.data_ptr<scalar_t>(),
+        M,
+        N,
+        K,
+        mat1_strideM,
+        out_strideM,
+        block_size_N,
+        block_size_K,
+        size_per_thread);
+  });
+
+  return out;
+}

csrc/cpu/sgl-kernels/gemm_int8.cpp

@@ -0,0 +1,440 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+namespace {
+
+template <typename scalar_t, bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, scalar_t* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn<at::BFloat16, has_bias, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, at::BFloat16* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+    static_assert(COLS % 2 == 0);
+
+    // prefetch distance
+    constexpr int PREFETCH_SIZE_K = 0;
+
+    __m512i va;
+    __m512i vb[COLS];
+    __m512i vc[ROWS * COLS];
+    __m512i vcomp[COLS];
+    __m512  vd0;
+    __m512  vd1[COLS];
+
+    // oops! 4x4 spills but luckly we use 4x2
+    __m512 vbias[COLS];
+
+    // [NOTE]: s8s8 igemm compensation in avx512-vnni
+    //
+    // avx512-vnni has no s8s8, so we need to change s8s8 to u8s8 with compensate:
+    //
+    //   a * b = (a + 128) * b - 128 * b
+    //   s   s       u       s    u    s
+    //
+    // 1) 128 * b is pre-computed when packing B to vnni formats
+    // 2) a + 128 is fused when dynamically quantize A
+    //
+    auto loadc = [&](auto i) {
+      vc[i] = _mm512_set1_epi32(0);
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K4 = K >> 2;
+    const int64_t lda4 = lda >> 2;
+    const int64_t ldb4 = ldb; // ldb * 4 >> 2;
+    const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A);
+    const int32_t* b_ptr = reinterpret_cast<const int32_t*>(B);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = _mm512_set1_epi32(a_ptr[row * lda4 + k]);
+      }
+      if constexpr (row == 0) {
+        vb[col] = _mm512_loadu_si512(b_ptr + k * ldb4 + col * 16);
+        if constexpr (PREFETCH_SIZE_K > 0) {
+          _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb4 + col * 16, _MM_HINT_T0);
+        }
+      }
+      vc[i] = _mm512_dpbusd_epi32(vc[i], va, vb[col]);
+    };
+    for (int64_t k = 0; k < K4; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      // load a scale
+      if constexpr(col == 0) {
+        vd0 = _mm512_set1_ps(As[row]);
+      }
+      // load b scale and vcomp per 2 vectors
+      // also load bias if any
+      if constexpr (row == 0) {
+        if constexpr (col % 2 == 0) {
+          vd1[col + 0] = _mm512_loadu_ps(Bs + col * 16);
+          vd1[col + 1] = _mm512_loadu_ps(Bs + col * 16 + 16);
+          vcomp[col + 0] = _mm512_loadu_si512(Bcomp + col * 16);
+          vcomp[col + 1] = _mm512_loadu_si512(Bcomp + col * 16 + 16);
+          if constexpr (has_bias) {
+            vbias[col + 0] = _mm512_loadu_ps(bias + col * 16);
+            vbias[col + 1] = _mm512_loadu_ps(bias + col * 16 + 16);
+          }
+        }
+      }
+
+      // for COLS = 2, 4 use 512bit store
+      if constexpr (col % 2 == 0) {
+        __m512 vc0 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[row * COLS + col + 0], vcomp[col + 0]));
+        __m512 vc1 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[row * COLS + col + 1], vcomp[col + 1]));
+        if constexpr (has_bias) {
+          vc0 = _mm512_fmadd_ps(_mm512_mul_ps(vc0, vd0), vd1[col + 0], vbias[col + 0]);
+          vc1 = _mm512_fmadd_ps(_mm512_mul_ps(vc1, vd0), vd1[col + 1], vbias[col + 1]);
+        } else {
+          vc0 = _mm512_mul_ps(_mm512_mul_ps(vc0, vd0), vd1[col + 0]);
+          vc1 = _mm512_mul_ps(_mm512_mul_ps(vc1, vd0), vd1[col + 1]);
+        }
+
+        _mm512_storeu_si512(
+            reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+            (__m512i)(_mm512_cvtne2ps_pbh(vc1, vc0)));
+      }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE)                          \
+    tinygemm_kernel_nn<scalar_t, has_bias, MB_SIZE, NB_SIZE>::apply(         \
+        A + mb_start * lda, B + nb_start * 4, C + mb_start * ldc + nb_start, \
+        As + mb_start, Bs + nb_start, Bcomp + nb_start,                      \
+        has_bias ? bias + nb_start : nullptr, K, lda, ldb, ldc);
+
+template <typename scalar_t, bool has_bias>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    int32_t* __restrict__ Ctmp,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg) {
+
+  // B compensation
+  const int32_t* Bcomp = reinterpret_cast<const int32_t*>(B + block_size_n() * K);
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int64_t mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        // mb_size = 1
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_NN(1, 32); break;
+        case 0x14: LAUNCH_TINYGEMM_KERNEL_NN(1, 64); break;
+        // mb_size = 2
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_NN(2, 32); break;
+        case 0x24: LAUNCH_TINYGEMM_KERNEL_NN(2, 64); break;
+        // mb_size = 3
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_NN(3, 32); break;
+        case 0x34: LAUNCH_TINYGEMM_KERNEL_NN(3, 64); break;
+        // mb_size = 4
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_NN(4, 32); break;
+        case 0x44: LAUNCH_TINYGEMM_KERNEL_NN(4, 64); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+template<typename scalar_t>
+void int8_scaled_mm_kernel_impl(
+    scalar_t* __restrict__ out,
+    const uint8_t* __restrict__ mat1,
+    const int8_t* __restrict__ mat2,
+    const float* __restrict__ scales1,
+    const float* __restrict__ scales2,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  // TODO: brgemm u8s8 depends on PyTorch 2.7 release.
+  const bool use_brgemm = false;
+
+  // K + 4 after compensation
+  const int64_t packed_row_size = get_row_size<int8_t>(K);
+
+  AT_DISPATCH_BOOL(bias != nullptr, has_bias, [&] {
+    at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t mb{0}, nb{0};
+      data_index_init(begin, mb, MB, nb, NB);
+
+      // for brgemm, use int32_t for accumulate
+      alignas(64) int32_t Ctmp[BLOCK_M * BLOCK_N];
+
+      for (int i = begin; i < end; ++i) {
+        UNUSED(i);
+        int mb_start = mb * BLOCK_M;
+        int mb_size = std::min(M - mb_start, BLOCK_M);
+        int nb_start = nb * BLOCK_N;
+        int nb_size = std::min(N - nb_start, BLOCK_N);
+
+        tinygemm_kernel<scalar_t, has_bias>(
+            /*   A */ mat1 + mb_start * K,
+            /*   B */ mat2 + nb_start * packed_row_size /* nb * BLOCK_N * (K + 4) */,
+            /*   C */ out + mb_start * N + nb_start,
+            /* Ctmp*/ Ctmp,
+            /*  As */ scales1 + mb_start,
+            /*  Bs */ scales2 + nb_start,
+            /* bias*/ bias + nb_start,
+            /*   M */ mb_size,
+            /*   N */ nb_size,
+            /*   K */ K,
+            /* lda */ K,
+            /* ldb */ nb_size,
+            /* ldc */ N,
+            /* brg */ use_brgemm);
+
+        // move to the next index
+        data_index_step(mb, MB, nb, NB);
+      }
+
+      if (use_brgemm) {
+        at::native::cpublas::brgemm_release();
+      }
+    });
+  });
+}
+
+} // anonymous namespace
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(const uint8_t* __restrict__ A, const int8_t* __restrict__ B, scalar_t* __restrict__ C,
+    int32_t* __restrict__ Ctmp,  const float* __restrict__ As, const float* __restrict__ Bs,
+    int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc, bool brg) {
+  tinygemm_kernel<scalar_t, false>(A, B, C, Ctmp, As, Bs, nullptr, M, N, K, lda, ldb, ldc, brg);
+}
+
+#define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE)                                                     \
+    template void tinygemm_kernel<TYPE>(                                                        \
+        const uint8_t* __restrict__ A, const int8_t* __restrict__ B, TYPE* __restrict__ C,      \
+        int32_t* __restrict__ Ctmp, const float* __restrict__ As, const float* __restrict__ Bs, \
+        int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc, bool brg)
+
+INSTANTIATE_TINYGEMM_TEMPLATE(at::BFloat16);
+INSTANTIATE_TINYGEMM_TEMPLATE(at::Half);
+
+std::tuple<at::Tensor, at::Tensor> per_token_quant_int8_cpu(at::Tensor& A) {
+  RECORD_FUNCTION("sgl-kernel::per_token_quant_int8_cpu", std::vector<c10::IValue>({A}));
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(A);
+  CHECK_DIM(2, A);
+
+  int64_t M = A.size(0);
+  int64_t K = A.size(1);
+  int64_t lda = A.stride(0);
+
+  const auto st = A.scalar_type();
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf,
+      "per_token_quant_int8: expect A to be bfloat16 or half.");
+
+  auto Aq = at::empty({M, K}, A.options().dtype(at::kByte));
+  auto As = at::empty({M}, A.options().dtype(at::kFloat));
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(st, "per_token_quant_int8", [&] {
+    uint8_t* __restrict__ Aq_data = Aq.data_ptr<uint8_t>();
+    float* __restrict__ As_data = As.data_ptr<float>();
+    const scalar_t* __restrict__ A_data = A.data_ptr<scalar_t>();
+
+    at::parallel_for(0, M, 0, [&] (int64_t begin, int64_t end) {
+      for (int64_t m = begin; m < end; ++m) {
+        quantize_row_int8<scalar_t>(
+            Aq_data + m * K,
+            As_data[m],
+            A_data + m * lda,
+            K);
+      }
+    });
+  });
+  return std::make_tuple(Aq, As);
+}
+
+// weight     :  static, per-channel, symmetric
+// activation : dynamic,   per-token, symmetric
+//
+// mat1    : [M, K]
+// mat2    : [N, K]
+// scales1 : [M]
+// scales2 : [N]
+// bias    : [N]
+// out     : [M, N]
+//
+at::Tensor int8_scaled_mm_cpu(at::Tensor& mat1, at::Tensor& mat2,
+    at::Tensor& scales1, at::Tensor& scales2,
+    std::optional<at::Tensor>& bias, at::ScalarType out_dtype, bool is_vnni) {
+  RECORD_FUNCTION("sgl-kernel::int8_scaled_mm_cpu", std::vector<c10::IValue>({mat1, mat2, scales1, scales2, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_INPUT(mat1);
+  CHECK_INPUT(mat2);
+  CHECK_INPUT(scales1);
+  CHECK_INPUT(scales2);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat1.size(1);
+
+  // see [NOTE]: s8s8 igemm compensation in avx512-vnni
+  CHECK_EQ(mat2.size(1), (int64_t)(is_vnni ? K + sizeof(int32_t) : K));
+  CHECK_EQ(scales1.numel(), M);
+  CHECK_EQ(scales2.numel(), N);
+
+  TORCH_CHECK(mat1.scalar_type() == at::kByte, "int8_scaled_mm: expect mat1 to be uint8.");
+  TORCH_CHECK(mat2.scalar_type() == at::kChar, "int8_scaled_mm: expect mat2 to be int8.");
+  TORCH_CHECK(scales1.scalar_type() == at::kFloat && scales2.scalar_type() == at::kFloat,
+      "int8_scaled_mm: expect scales to be float32.");
+
+  auto out = at::empty({M, N}, mat1.options().dtype(out_dtype));
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "int8_scaled_mm_kernel_impl", [&] {
+    int8_scaled_mm_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        mat1.data_ptr<uint8_t>(),
+        packed_w.data_ptr<int8_t>(),
+        scales1.data_ptr<float>(),
+        scales2.data_ptr<float>(),
+        bias_data,
+        M,
+        N,
+        K);
+  });
+  return out;
+}
+
+// fused `per_token_quant_int8_cpu` and `int8_scaled_mm_cpu`
+at::Tensor int8_scaled_mm_with_quant(at::Tensor& mat1, at::Tensor& mat2, at::Tensor& scales2,
+    const std::optional<at::Tensor>& bias, at::ScalarType out_dtype, bool is_vnni) {
+  RECORD_FUNCTION("sgl-kernel::int8_scaled_mm_cpu", std::vector<c10::IValue>({mat1, mat2, scales2, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(mat1);
+  CHECK_INPUT(mat2);
+  CHECK_INPUT(scales2);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat1.size(1);
+  int64_t lda = mat1.stride(0);
+
+  // see [NOTE]: s8s8 igemm compensation in avx512-vnni
+  CHECK_EQ(mat2.size(1), (int64_t)(is_vnni ? K + sizeof(int32_t) : K));
+  CHECK_EQ(scales2.numel(), N);
+
+  const auto st = mat1.scalar_type();
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf,
+      "int8_scaled_mm_with_quant: expect A to be bfloat16 or half.");
+  TORCH_CHECK(st == out_dtype,
+      "int8_scaled_mm_with_quant: expect A has same dtype with out_dtype.");
+  TORCH_CHECK(mat2.scalar_type() == at::kChar,
+      "int8_scaled_mm_with_quant: expect mat2 to be int8.");
+  TORCH_CHECK(scales2.scalar_type() == at::kFloat,
+      "int8_scaled_mm_with_quant: expect scales to be float32.");
+
+  const int64_t buffer_size = M * K + M * sizeof(float);
+  auto buffer = at::empty({buffer_size}, mat1.options().dtype(at::kByte));
+  auto out = at::empty({M, N}, mat1.options().dtype(out_dtype));
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "int8_scaled_mm_with_quant_kernel_impl", [&] {
+    uint8_t* __restrict__ Aq_data = buffer.data_ptr<uint8_t>();
+    float* __restrict__ As_data = (float*)((void*)(Aq_data + M * K));
+    const scalar_t* __restrict__ A_data = mat1.data_ptr<scalar_t>();
+
+    at::parallel_for(0, M, 0, [&] (int64_t begin, int64_t end) {
+      for (int64_t m = begin; m < end; ++m) {
+        quantize_row_int8<scalar_t>(
+            Aq_data + m * K,
+            As_data[m],
+            A_data + m * lda,
+            K);
+      }
+    });
+
+    int8_scaled_mm_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        Aq_data,
+        packed_w.data_ptr<int8_t>(),
+        As_data,
+        scales2.data_ptr<float>(),
+        bias_data,
+        M,
+        N,
+        K);
+  });
+  return out;
+}

csrc/cpu/sgl-kernels/moe_fp8.cpp

@@ -0,0 +1,502 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "gemm.h"
+#include "vec.h"
+
+// clang-format off
+
+namespace {
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
+  using Vec = at::vec::Vectorized<scalar_t>;
+  // no remainder
+  #pragma GCC unroll 4
+  for (int64_t d = 0; d < size; d += Vec::size()) {
+    Vec data = Vec::loadu(input + d);
+    data.store(out + d);
+  }
+}
+
+template <typename scalar_t>
+inline void copy_mul_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, float weight, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec weight_vec = fVec(weight);
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    bVec x = bVec::loadu(input + d);
+    fVec x0, x1;
+    std::tie(x0, x1) = at::vec::convert_to_float(x);
+    x0 = x0 * weight_vec;
+    x1 = x1 * weight_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] * weight);
+  }
+}
+
+// acc from [topk, K] to [K]
+template <typename scalar_t>
+inline void sum_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t topk, int64_t K) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  if (topk == 1) {
+    // do copy for topk = 1
+    copy_stub(out, input, K);
+  } else {
+    // do sum for topk != 1
+    int64_t d;
+    #pragma GCC unroll 4
+    for (d = 0; d <= K - kVecSize; d += kVecSize) {
+      fVec sum_fvec0 = fVec(0.f);
+      fVec sum_fvec1 = fVec(0.f);
+      for (int t = 0; t < topk; ++t) {
+        bVec x_bvec = bVec::loadu(input + t * K + d);
+        fVec x_fvec0, x_fvec1;
+        std::tie(x_fvec0, x_fvec1) = at::vec::convert_to_float(x_bvec);
+
+        sum_fvec0 += x_fvec0;
+        sum_fvec1 += x_fvec1;
+      }
+      bVec out_bvec = convert_from_float_ext<scalar_t>(sum_fvec0, sum_fvec1);
+      out_bvec.store(out + d);
+    }
+    for (; d < K; ++d) {
+      float sum_val = 0.f;
+      for (int t = 0; t < topk; ++t) {
+        sum_val += static_cast<float>(input[t * K + d]);
+      }
+      out[d] = static_cast<scalar_t>(sum_val);
+    }
+  }
+}
+
+// out = input + input2 * scale
+template <typename scalar_t>
+inline void add_mul_stub(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ input,
+    const scalar_t* __restrict__ input2,
+    float scale,
+    int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec s_vec = fVec(scale);
+
+  int64_t d;
+#pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    bVec x_bvec = bVec::loadu(input + d);
+    fVec x0, x1;
+    std::tie(x0, x1) = at::vec::convert_to_float(x_bvec);
+
+    bVec y_bvec = bVec::loadu(input2 + d);
+    fVec y0, y1;
+    std::tie(y0, y1) = at::vec::convert_to_float(y_bvec);
+
+    x0 = x0 + y0 * s_vec;
+    x1 = x1 + y1 * s_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + float(input2[d]) * scale);
+  }
+}
+
+template <typename scalar_t>
+inline void silu_and_mul_stub(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ input,
+    const scalar_t* __restrict__ input2,
+    int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  const fVec one = fVec(1.f);
+
+  // no remainder
+#pragma GCC unroll 4
+  for (int64_t d = 0; d < size; d += bVec::size()) {
+    bVec x = bVec::loadu(input + d);
+    fVec x0, x1;
+    std::tie(x0, x1) = at::vec::convert_to_float(x);
+    bVec y = bVec::loadu(input2 + d);
+    fVec y0, y1;
+    std::tie(y0, y1) = at::vec::convert_to_float(y);
+    x0 = x0 / (one + x0.neg().exp_u20());
+    x1 = x1 / (one + x1.neg().exp_u20());
+    x0 = x0 * y0;
+    x1 = x1 * y1;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+}
+
+} // anonymous namespace
+
+template <typename scalar_t>
+void fused_experts_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    scalar_t* __restrict__ A_tmp,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 1: intermediate_cache0 = hidden_states @ w1
+  const int64_t MB = div_up(num_tokens_post_pad, BLOCK_M);
+  const int64_t NB = div_up(2 * N, BLOCK_N);
+  int64_t scale_size_N = div_up(2 * N, block_size_N);
+  int64_t scale_size_K = div_up(K, block_size_K);
+  int64_t blocks_n_per_group = block_size_N / BLOCK_N;
+
+  const int64_t stride_e = 2 * N * K;
+  const int64_t stride_n = K;
+
+  // here we only parallel on half of 2N to fuse silu_and_mul with gemm
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    scalar_t* __restrict__ A = A_tmp + tid * BLOCK_M * K;
+
+    bool is_brgemm_used = false;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+
+      int64_t n_size = std::min(2 * N - nb * BLOCK_N, BLOCK_N);
+
+      // B shape [K, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const at::Float8_e4m3fn* __restrict__ B = packed_w1 + expert_id * stride_e + nb * BLOCK_N * stride_n;
+      const float* __restrict__ Bs = w1s + expert_id * scale_size_N * scale_size_K + (nb / blocks_n_per_group) * scale_size_K;
+
+      // 1.a load A
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+
+      const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(m_size);
+      is_brgemm_used = is_brgemm_used || use_brgemm;
+
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m] / topk;
+        copy_stub(A + m * K, input + index * K, K);
+      }
+
+      const int64_t offset = offsets[mb];
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ A,
+          /*   B            */ B,
+          /*   C            */ ic0 + offset * 2 * N + nb * BLOCK_N,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ Bs,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ K,
+          /*   lda          */ K,
+          /*   ldb          */ n_size,
+          /*   ldc          */ 2 * N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+    }
+
+    if (is_brgemm_used) {
+      at::native::cpublas::brgemm_release();
+    }
+  });
+
+  // stage 1.5: intermediate_cache1 = silu(intermediate_cache0)
+  at::parallel_for(0, M * topk, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      silu_and_mul_stub(
+          ic1 + m * N,
+          ic0 + m * 2 * N,
+          ic0 + m * 2 * N + N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [E, K, N] as [E, OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(OC, BLOCK_N);
+  scale_size_N = div_up(K, block_size_N);
+  scale_size_K = div_up(N, block_size_K);
+  const int64_t stride_e2 = OC * IC;
+  const int64_t stride_oc = IC;
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    int tid = at::get_thread_num();
+    alignas(64) scalar_t C[BLOCK_M * BLOCK_K];
+
+    bool is_brgemm_used = false;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(m_size);
+      is_brgemm_used = is_brgemm_used || use_brgemm;
+
+      // A ptr from ic1 of [M * topk, N] in sorted order
+      // so as to avoid copy A to tmp buffer again
+      const scalar_t* __restrict__ A = ic1 + offsets[mb] * N;
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+
+      // B shape [IC, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const at::Float8_e4m3fn* __restrict__ B = packed_w2 + expert_id * stride_e2 + nb * BLOCK_N * stride_oc;
+      const float* __restrict__ Bs = w2s + expert_id * scale_size_N * scale_size_K + (nb / blocks_n_per_group) * scale_size_K;
+
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ A,
+          /*   B            */ B,
+          /*   C            */ C,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ Bs,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ IC,
+          /*   lda          */ IC,
+          /*   ldb          */ n_size,
+          /*   ldc          */ BLOCK_N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+
+      // 2.b copy from C to ic2 in original order
+      //   and also mul topk_weights in float32
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m];
+        float weight = topk_weights[index];
+        copy_mul_stub(ic2 + index * K + nb * BLOCK_N, C + m * BLOCK_N, weight, n_size);
+      }
+    }
+
+    if (is_brgemm_used) {
+      at::native::cpublas::brgemm_release();
+    }
+  });
+
+  // stage 3: out = intermediate_cache2.sum(dim=1)
+  //   from [M, topk, K] to [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      sum_stub(output + m * K, ic2 + m * topk * K, topk, K);
+    }
+  });
+}
+
+#define INSTANTIATE_MOE_FP8_TEMPLATE(TYPE)             \
+  template void fused_experts_fp8_kernel_impl<TYPE>(   \
+      TYPE* __restrict__ output,                       \
+      TYPE* __restrict__ ic0,                          \
+      TYPE* __restrict__ ic1,                          \
+      TYPE* __restrict__ ic2,                          \
+      TYPE* __restrict__ A_tmp,                        \
+      TYPE* __restrict__ B_tmp,                        \
+      float* __restrict__ C_tmp,                       \
+      const TYPE* __restrict__ input,                  \
+      const at::Float8_e4m3fn* __restrict__ packed_w1, \
+      const at::Float8_e4m3fn* __restrict__ packed_w2, \
+      const float* __restrict__ w1s,                   \
+      const float* __restrict__ w2s,                   \
+      int64_t block_size_N,                            \
+      int64_t block_size_K,                            \
+      const float* __restrict__ topk_weights,          \
+      const int32_t* __restrict__ sorted_ids,          \
+      const int32_t* __restrict__ expert_ids,          \
+      const int32_t* __restrict__ offsets,             \
+      int64_t M,                                       \
+      int64_t N,                                       \
+      int64_t K,                                       \
+      int64_t E,                                       \
+      int64_t topk,                                    \
+      int64_t num_tokens_post_pad)
+
+INSTANTIATE_MOE_FP8_TEMPLATE(at::BFloat16);
+INSTANTIATE_MOE_FP8_TEMPLATE(at::Half);
+
+template <typename scalar_t>
+void shared_expert_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 1: intermediate_cache0 = hidden_states @ w1
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(2 * N, BLOCK_N);
+  int64_t scale_size_K = div_up(K, block_size_K);
+  int64_t blocks_n_per_group = block_size_N / BLOCK_N;
+
+  const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(M);
+
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    int tid = at::get_thread_num();
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+      int64_t n_size = std::min(2 * N - nb * BLOCK_N, BLOCK_N);
+
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ input + mb * BLOCK_M * K,
+          /*   B            */ packed_w1 + nb * BLOCK_N * K,
+          /*   C            */ ic0 + mb * BLOCK_M * 2 * N + nb * BLOCK_N,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ w1s + (nb / blocks_n_per_group) * scale_size_K,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ K,
+          /*   lda          */ K,
+          /*   ldb          */ n_size,
+          /*   ldc          */ 2 * N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+    }
+
+    if (use_brgemm) {
+      at::native::cpublas::brgemm_release();
+    }
+  });
+
+  // stage 1.5: intermediate_cache1 = silu(intermediate_cache0)
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      silu_and_mul_stub(
+          ic1 + m * N,
+          ic0 + m * 2 * N,
+          ic0 + m * 2 * N + N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [K, N] as [OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(K, BLOCK_N);
+  scale_size_K = div_up(N, block_size_K);
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    int tid = at::get_thread_num();
+    alignas(64) scalar_t C[BLOCK_M * BLOCK_K];
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      // 2.a gemm: C = A @ B
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ ic1 + mb * BLOCK_M * N,
+          /*   B            */ packed_w2 + nb * BLOCK_N * N,
+          /*   C            */ C,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ w2s + (nb / blocks_n_per_group) * scale_size_K,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ IC,
+          /*   lda          */ IC,
+          /*   ldb          */ n_size,
+          /*   ldc          */ BLOCK_N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+
+      // 2.b copy from C to output and add fused_experts_out
+      scalar_t* __restrict__ out = output + mb * BLOCK_M * K + nb * BLOCK_N;
+      const scalar_t* __restrict__ fused_out = fused_experts_out + mb * BLOCK_M * K + nb * BLOCK_N;
+      for (int64_t m = 0; m < m_size; ++m) {
+        add_mul_stub(out + m * K, C + m * BLOCK_N, fused_out + m * K, routed_scaling_factor, n_size);
+      }
+    }
+  });
+
+  if (use_brgemm) {
+    at::native::cpublas::brgemm_release();
+  }
+}
+
+#define INSTANTIATE_SHARED_EXPERT_FP8_TEMPLATE(TYPE)   \
+  template void shared_expert_fp8_kernel_impl<TYPE>(   \
+      TYPE* __restrict__ output,                       \
+      TYPE* __restrict__ ic0,                          \
+      TYPE* __restrict__ ic1,                          \
+      TYPE* __restrict__ B_tmp,                        \
+      float* __restrict__ C_tmp,                       \
+      const TYPE* __restrict__ input,                  \
+      const at::Float8_e4m3fn* __restrict__ packed_w1, \
+      const at::Float8_e4m3fn* __restrict__ packed_w2, \
+      const float* __restrict__ w1s,                   \
+      const float* __restrict__ w2s,                   \
+      int64_t block_size_N,                            \
+      int64_t block_size_K,                            \
+      const TYPE* __restrict__ fused_experts_out,      \
+      float routed_scaling_factor,                     \
+      int64_t M,                                       \
+      int64_t N,                                       \
+      int64_t K)
+
+INSTANTIATE_SHARED_EXPERT_FP8_TEMPLATE(at::BFloat16);
+INSTANTIATE_SHARED_EXPERT_FP8_TEMPLATE(at::Half);

csrc/cpu/sgl-kernels/moe_int8.cpp

@@ -0,0 +1,769 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+namespace {
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
+  using Vec = at::vec::Vectorized<scalar_t>;
+  // no remainder
+  #pragma GCC unroll 4
+  for (int64_t d = 0; d < size; d += Vec::size()) {
+    Vec data = Vec::loadu(input + d);
+    data.store(out + d);
+  }
+}
+
+template <>
+inline void copy_stub<uint8_t>(uint8_t* __restrict__ out, const uint8_t* __restrict__ input, int64_t size) {
+  // size might be 64x + 32
+  std::memcpy(out, input, size * sizeof(uint8_t));
+}
+
+template <typename scalar_t>
+inline void copy_mul_stub(scalar_t* __restrict__ out, const float* __restrict__ input, float weight, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec weight_vec = fVec(weight);
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d) * weight_vec;
+    fVec data1 = fVec::loadu(input + d + fVec::size()) * weight_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] * weight);
+  }
+}
+
+// acc from [topk, K] to [K]
+template <typename scalar_t>
+inline void sum_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t topk, int64_t K) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  if (topk == 1) {
+    // do copy for topk = 1
+    copy_stub(out, input, K);
+  } else {
+    // do sum for topk != 1
+    int64_t d;
+    #pragma GCC unroll 4
+    for (d = 0; d <= K - kVecSize; d += kVecSize) {
+      fVec sum_fvec0 = fVec(0.f);
+      fVec sum_fvec1 = fVec(0.f);
+      for (int t = 0; t < topk; ++t) {
+        bVec x_bvec = bVec::loadu(input + t * K + d);
+        fVec x_fvec0, x_fvec1;
+        std::tie(x_fvec0, x_fvec1) = at::vec::convert_to_float(x_bvec);
+
+        sum_fvec0 += x_fvec0;
+        sum_fvec1 += x_fvec1;
+      }
+      bVec out_bvec = convert_from_float_ext<scalar_t>(sum_fvec0, sum_fvec1);
+      out_bvec.store(out + d);
+    }
+    for (; d < K; ++d) {
+      float sum_val = 0.f;
+      for (int t = 0; t < topk; ++t) {
+        sum_val += static_cast<float>(input[t * K + d]);
+      }
+      out[d] = static_cast<scalar_t>(sum_val);
+    }
+  }
+}
+
+// out = input + input2 * scale
+template <typename scalar_t>
+inline void add_mul_stub(scalar_t* __restrict__ out, const float* __restrict__ input,
+    const scalar_t* __restrict__ input2, float scale, int64_t size) {
+
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec s_vec = fVec(scale);
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec x0 = fVec::loadu(input + d);
+    fVec x1 = fVec::loadu(input + d + fVec::size());
+
+    bVec y_bvec = bVec::loadu(input2 + d);
+    fVec y0, y1;
+    std::tie(y0, y1) = at::vec::convert_to_float(y_bvec);
+
+    x0 = x0 + y0 * s_vec;
+    x1 = x1 + y1 * s_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + float(input2[d]) * scale);
+  }
+}
+
+/// gemm for w13
+template <typename scalar_t, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B0, const int8_t* __restrict__ B1, scalar_t* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs0, const float* __restrict__ Bs1,
+      const int32_t* __restrict__ Bcomp0, const int32_t* __restrict__ Bcomp1,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni<at::BFloat16, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B0, const int8_t* __restrict__ B1, at::BFloat16* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs0, const float* __restrict__ Bs1,
+      const int32_t* __restrict__ Bcomp0, const int32_t* __restrict__ Bcomp1,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+    static_assert(COLS % 2 == 0);
+
+    __m512i va;
+    __m512i vb0[COLS];
+    __m512i vb1[COLS];
+    __m512i vc0[ROWS * COLS];
+    __m512i vc1[ROWS * COLS];
+    __m512i vcomp0[COLS];
+    __m512i vcomp1[COLS];
+    __m512  was;
+    __m512  vbs0[COLS];
+    __m512  vbs1[COLS];
+
+    auto loadc = [&](auto i) {
+      vc0[i] = _mm512_set1_epi32(0);
+      vc1[i] = _mm512_set1_epi32(0);
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K4 = K >> 2;
+    const int64_t lda4 = lda >> 2;
+    const int64_t ldb4 = ldb; // ldb * 4 >> 2;
+    const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A);
+    const int32_t* b0_ptr = reinterpret_cast<const int32_t*>(B0);
+    const int32_t* b1_ptr = reinterpret_cast<const int32_t*>(B1);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = _mm512_set1_epi32(a_ptr[row * lda4 + k]);
+      }
+      if constexpr (row == 0) {
+        vb0[col] = _mm512_loadu_si512(b0_ptr + k * ldb4 + col * 16);
+        vb1[col] = _mm512_loadu_si512(b1_ptr + k * ldb4 + col * 16);
+      }
+      vc0[i] = _mm512_dpbusd_epi32(vc0[i], va, vb0[col]);
+      vc1[i] = _mm512_dpbusd_epi32(vc1[i], va, vb1[col]);
+    };
+    for (int64_t k = 0; k < K4; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto scalec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      // load a scale
+      if constexpr(col == 0) {
+        was = _mm512_set1_ps(As[row]);
+      }
+      // load b scale and vcomp
+      if constexpr (row == 0) {
+        vbs0[col] = _mm512_loadu_ps(Bs0 + col * 16);
+        vbs1[col] = _mm512_loadu_ps(Bs1 + col * 16);
+        vcomp0[col] = _mm512_loadu_si512(Bcomp0 + col * 16);
+        vcomp1[col] = _mm512_loadu_si512(Bcomp1 + col * 16);
+      }
+      __m512 c0 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc0[i], vcomp0[col]));
+      __m512 c1 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc1[i], vcomp1[col]));
+      vc0[i] = _mm512_castps_si512(_mm512_mul_ps(_mm512_mul_ps(c0, was), vbs0[col]));
+      vc1[i] = _mm512_castps_si512(_mm512_mul_ps(_mm512_mul_ps(c1, was), vbs1[col]));
+    };
+    Unroll<ROWS * COLS>{}(scalec);
+
+    using Vec = at::vec::Vectorized<float>;
+    const Vec one = Vec(1.f);
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+      // for COLS = 2, 4 use 512bit store
+      if constexpr (col % 2 == 0) {
+        Vec x0 = _mm512_castsi512_ps(vc0[row * COLS + col + 0]);
+        Vec x1 = _mm512_castsi512_ps(vc0[row * COLS + col + 1]);
+        Vec y0 = _mm512_castsi512_ps(vc1[row * COLS + col + 0]);
+        Vec y1 = _mm512_castsi512_ps(vc1[row * COLS + col + 1]);
+        // silu
+        x0 = x0 / (one + x0.neg().exp_u20());
+        x1 = x1 / (one + x1.neg().exp_u20());
+        // mul
+        x0 = x0 * y0;
+        x1 = x1 * y1;
+
+        _mm512_storeu_si512(
+            reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+            (__m512i)(_mm512_cvtne2ps_pbh(__m512(x1), __m512(x0))));
+        }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_VNNI(MB_SIZE, NB_SIZE)                        \
+    tinygemm_kernel_vnni<scalar_t, MB_SIZE, NB_SIZE>::apply(                 \
+        A + mb_start * lda, B0 + nb_start * 4, B1 + nb_start * 4,            \
+        C + mb_start * ldc + nb_start, As + mb_start,                        \
+        Bs0 + nb_start, Bs1 + nb_start, Bcomp0 + nb_start, Bcomp1 + nb_start,\
+        K, lda, ldb, ldc);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B0,
+    const int8_t* __restrict__ B1,
+    scalar_t* __restrict__ C,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs0,
+    const float* __restrict__ Bs1,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc) {
+
+  const int32_t* Bcomp0 = reinterpret_cast<const int32_t*>(B0 + block_size_n() * K);
+  const int32_t* Bcomp1 = reinterpret_cast<const int32_t*>(B1 + block_size_n() * K);
+
+  // pattern: 1-(2+2)-(8+8)
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 32;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_VNNI(1, 32); break;
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_VNNI(2, 32); break;
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_VNNI(3, 32); break;
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_VNNI(4, 32); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+/// gemm for w2
+template <typename scalar_t, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni2 {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, float* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni2<at::BFloat16, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, float* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+    static_assert(COLS % 2 == 0);
+
+    __m512i va;
+    __m512i vb[COLS];
+    __m512i vc[ROWS * COLS];
+    __m512i vcomp[COLS];
+    __m512  was;
+    __m512  vbs[COLS];
+
+    auto loadc = [&](auto i) {
+      vc[i] = _mm512_set1_epi32(0);
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K4 = K >> 2;
+    const int64_t lda4 = lda >> 2;
+    const int64_t ldb4 = ldb; // ldb * 4 >> 2;
+    const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A);
+    const int32_t* b_ptr = reinterpret_cast<const int32_t*>(B);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = _mm512_set1_epi32(a_ptr[row * lda4 + k]);
+      }
+      if constexpr (row == 0) {
+        vb[col] = _mm512_loadu_si512(b_ptr + k * ldb4 + col * 16);
+      }
+      vc[i] = _mm512_dpbusd_epi32(vc[i], va, vb[col]);
+    };
+    for (int64_t k = 0; k < K4; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      // load a scale
+      if constexpr(col == 0) {
+        was = _mm512_set1_ps(As[row]);
+      }
+      // load b scale and vcomp per 2 vectors
+      // also load bias if any
+      if constexpr (row == 0) {
+        if constexpr (col % 2 == 0) {
+          vbs[col + 0] = _mm512_loadu_ps(Bs + col * 16);
+          vbs[col + 1] = _mm512_loadu_ps(Bs + col * 16 + 16);
+          vcomp[col + 0] = _mm512_loadu_si512(Bcomp + col * 16);
+          vcomp[col + 1] = _mm512_loadu_si512(Bcomp + col * 16 + 16);
+        }
+      }
+      __m512 x = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[i], vcomp[col]));
+      x = _mm512_mul_ps(_mm512_mul_ps(x, was), vbs[col]);
+      _mm512_storeu_ps(reinterpret_cast<__m512*>(C + row * ldc + col * 16), x);
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_VNNI2(MB_SIZE, NB_SIZE)                       \
+    tinygemm_kernel_vnni2<scalar_t, MB_SIZE, NB_SIZE>::apply(                \
+        A + mb_start * lda, B + nb_start * 4, C + mb_start * ldc + nb_start, \
+        As + mb_start, Bs + nb_start, Bcomp + nb_start,                      \
+        K, lda, ldb, ldc);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B,
+    float* __restrict__ C,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc) {
+
+  // B compensation
+  const int32_t* Bcomp = reinterpret_cast<const int32_t*>(B + block_size_n() * K);
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int64_t mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_VNNI2(1, 32); break;
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_VNNI2(2, 32); break;
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_VNNI2(3, 32); break;
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_VNNI2(4, 32); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+} // anonymous namespace
+
+template <typename scalar_t>
+void fused_experts_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    uint8_t* __restrict__ A_tmp,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad) {
+
+  // handle 2 tiles per block
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 0: quantize input to uint8, [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * K,
+          As_tmp[m],
+          input + m * K,
+          K);
+    }
+  });
+
+  // stage 1: intermediate_cache1 = silu(hidden_states @ w1)
+  const int64_t MB = div_up(num_tokens_post_pad, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  // strides for w1: [E, 2N, K]
+  TORCH_CHECK(N % BLOCK_N == 0, "Fixme when N is not multiples of ", BLOCK_N);
+
+  // K and N are packed for int8
+  const int64_t packed_K = get_row_size<int8_t>(K);
+  const int64_t packed_N = get_row_size<int8_t>(N);
+
+  const int64_t stride_e = 2 * N * packed_K;
+  const int64_t stride_n = packed_K;
+  // here we only parallel on half of 2N to fuse silu_and_mul with gemm
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    uint8_t* __restrict__ A = A_tmp + tid * BLOCK_M * K;
+
+    alignas(64) float As[BLOCK_M];
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+
+      // nb0 from top half and nb1 from bottom half
+      int64_t nb0 = nb, nb1 = nb + NB;
+      int64_t n_size = std::min(N - nb0 * BLOCK_N, BLOCK_N);
+
+      // B shape [K, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const int8_t* __restrict__ B0 = packed_w1 + expert_id * stride_e + nb0 * BLOCK_N * stride_n;
+      const int8_t* __restrict__ B1 = packed_w1 + expert_id * stride_e + nb1 * BLOCK_N * stride_n;
+      const float* __restrict__ Bs0 = w1s + expert_id * 2 * N + nb0 * BLOCK_N;
+      const float* __restrict__ Bs1 = w1s + expert_id * 2 * N + nb1 * BLOCK_N;
+
+      // 1.a load A
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m] / topk;
+        copy_stub(A + m * K, Aq_tmp + index * K, K);
+        As[m] = As_tmp[index];
+      }
+
+      // fused 1.b: silu_and_mul(A @ B0, A @ B1)
+      const int64_t offset = offsets[mb];
+      tinygemm_kernel(
+          /* A     */ A,
+          /* B0    */ B0,
+          /* B1    */ B1,
+          /* C     */ ic1 + offset * N + nb * BLOCK_N,
+          /* As    */ As,
+          /* Bs0   */ Bs0,
+          /* Bs1   */ Bs1,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ K,
+          /* lda   */ K,
+          /* ldb   */ n_size,
+          /* ldc   */ N);
+    }
+  });
+
+  // stage 1.5: quantize ic1 to uint8, [M * topk, N]
+  at::parallel_for(0, M * topk, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * N,
+          As_tmp[m],
+          ic1 + m * N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [E, K, N] as [E, OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(OC, BLOCK_N);
+  const int64_t stride_e2 = OC * packed_N;
+  const int64_t stride_oc = packed_N;
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    // we won't be using C1 for gemm2
+    float* __restrict__ C = C_tmp + tid * 2 * BLOCK_M * BLOCK_N;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      // A ptr from ic1 of [M * topk, N] in sorted order
+      // so as to avoid copy A to tmp buffer again
+      const uint8_t* __restrict__ A = Aq_tmp + offsets[mb] * N;
+      const float* __restrict__ As = As_tmp + offsets[mb];
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+
+      // B shape [IC, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const int8_t* __restrict__ B = packed_w2 + expert_id * stride_e2 + nb * BLOCK_N * stride_oc;
+      const float* __restrict__ Bs = w2s + expert_id * K + nb * BLOCK_N;
+
+      // 2.a gemm: C = A @ B
+      tinygemm_kernel<scalar_t>(
+          /* A     */ A,
+          /* B     */ B,
+          /* C     */ C,
+          /* As    */ As,
+          /* Bs    */ Bs,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ IC,
+          /* lda   */ IC,
+          /* ldb   */ n_size,
+          /* ldc   */ BLOCK_N);
+
+      // 2.b copy from C to ic2 in original order
+      //   and also mul topk_weights in float32
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m];
+        float weight = topk_weights[index];
+        copy_mul_stub(ic2 + index * K + nb * BLOCK_N, C + m * BLOCK_N, weight, n_size);
+      }
+    }
+  });
+
+  // stage 3: out = intermediate_cache2.sum(dim=1)
+  //   from [M, topk, K] to [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      sum_stub(output + m * K, ic2 + m * topk * K, topk, K);
+    }
+  });
+}
+
+#define INSTANTIATE_MOE_INT8_TEMPLATE(TYPE)                                                  \
+  template void fused_experts_int8_kernel_impl<TYPE> (                                       \
+      TYPE* __restrict__ output, TYPE* __restrict__ ic1,                                     \
+      TYPE* __restrict__ ic2, uint8_t* __restrict__ A_tmp,                                   \
+      float* __restrict__ C_tmp, uint8_t* __restrict__ Aq_tmp,                               \
+      float* __restrict__ As_tmp, const TYPE* __restrict__ input,                            \
+      const int8_t* __restrict__ packed_w1, const int8_t* __restrict__ packed_w2,            \
+      const float* __restrict__ w1s, const float* __restrict__ w2s,                          \
+      const float* __restrict__ topk_weights, const int32_t* __restrict__ sorted_ids,        \
+      const int32_t* __restrict__ expert_ids, const int32_t* __restrict__ offsets,           \
+      int64_t M, int64_t N, int64_t K, int64_t E, int64_t topk, int64_t num_tokens_post_pad)
+
+INSTANTIATE_MOE_INT8_TEMPLATE(at::BFloat16);
+INSTANTIATE_MOE_INT8_TEMPLATE(at::Half);
+
+template <typename scalar_t>
+void shared_expert_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K) {
+
+  // handle 2 tiles per block
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 0: quantize input to uint8, [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * K,
+          As_tmp[m],
+          input + m * K,
+          K);
+    }
+  });
+
+   // stage 1: intermediate_cache1 = silu(hidden_states @ w1)
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  TORCH_CHECK(N % BLOCK_N == 0, "Fixme when N is not multiples of ", BLOCK_N);
+
+  // K and N are packed for int8
+  const int64_t packed_K = get_row_size<int8_t>(K);
+  const int64_t packed_N = get_row_size<int8_t>(N);
+  const int64_t stride_n = packed_K;
+
+  // here we only parallel on half of 2N to fuse silu_and_mul with gemm
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+
+      // nb0 from top half and nb1 from bottom half
+      int64_t nb0 = nb, nb1 = nb + NB;
+      int64_t n_size = std::min(N - nb0 * BLOCK_N, BLOCK_N);
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+
+      // A shape [m_size, K]
+      const uint8_t* A = Aq_tmp + mb * BLOCK_M * K;
+      const float* As = As_tmp + mb * BLOCK_M;
+
+      // B shape [K, n_size] in vnni format
+      const int8_t* __restrict__ B0 = packed_w1 + nb0 * BLOCK_N * stride_n;
+      const int8_t* __restrict__ B1 = packed_w1 + nb1 * BLOCK_N * stride_n;
+      const float* __restrict__ Bs0 = w1s + nb0 * BLOCK_N;
+      const float* __restrict__ Bs1 = w1s + nb1 * BLOCK_N;
+
+      // fused 1.b: silu_and_mul(A @ B0, A @ B1)
+      tinygemm_kernel(
+          /* A     */ A,
+          /* B0    */ B0,
+          /* B1    */ B1,
+          /* C     */ ic1 + mb * BLOCK_M * N + nb * BLOCK_N,
+          /* As    */ As,
+          /* Bs0   */ Bs0,
+          /* Bs1   */ Bs1,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ K,
+          /* lda   */ K,
+          /* ldb   */ n_size,
+          /* ldc   */ N);
+    }
+  });
+
+  // stage 1.5: quantize ic1 to uint8, [M * topk, N]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * N,
+          As_tmp[m],
+          ic1 + m * N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [K, N] as [OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(OC, BLOCK_N);
+  const int64_t stride_oc = packed_N;
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    // we won't be using C1 for gemm2
+    float* __restrict__ C = C_tmp + tid * 2 * BLOCK_M * BLOCK_N;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      // A shape [m_size, IC]
+      const uint8_t* __restrict__ A = Aq_tmp + mb * BLOCK_M * N;
+      const float* __restrict__ As = As_tmp + mb * BLOCK_M;
+
+      // B shape [IC, n_size] in vnni format
+      const int8_t* __restrict__ B = packed_w2 + nb * BLOCK_N * stride_oc;
+      const float* __restrict__ Bs = w2s + nb * BLOCK_N;
+
+      // 2.a gemm: C = A @ B
+      tinygemm_kernel<scalar_t>(
+          /* A     */ A,
+          /* B     */ B,
+          /* C     */ C,
+          /* As    */ As,
+          /* Bs    */ Bs,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ IC,
+          /* lda   */ IC,
+          /* ldb   */ n_size,
+          /* ldc   */ BLOCK_N);
+
+      // 2.b copy from C to output and add fused_experts_out
+      scalar_t* __restrict__ out = output + mb * BLOCK_M * K + nb * BLOCK_N;
+      const scalar_t* __restrict__ fused_out = fused_experts_out + mb * BLOCK_M * K + nb * BLOCK_N;
+      for (int64_t m = 0; m < m_size; ++m) {
+        add_mul_stub(out + m * K, C + m * BLOCK_N, fused_out + m * K, routed_scaling_factor, n_size);
+      }
+    }
+  });
+}
+
+#define INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(TYPE)                                        \
+  template void shared_expert_int8_kernel_impl<TYPE> (                                       \
+      TYPE* __restrict__ output, TYPE* __restrict__ ic1,                                     \
+      float* __restrict__ C_tmp, uint8_t* __restrict__ Aq_tmp,                               \
+      float* __restrict__ As_tmp, const TYPE* __restrict__ input,                            \
+      const int8_t* __restrict__ packed_w1, const int8_t* __restrict__ packed_w2,            \
+      const float* __restrict__ w1s, const float* __restrict__ w2s,                          \
+      const TYPE* __restrict__ fused_experts_out, float routed_scaling_factor,               \
+      int64_t M, int64_t N, int64_t K)
+
+INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(at::BFloat16);
+INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(at::Half);

csrc/cpu/sgl-kernels/vec.h

@@ -0,0 +1,308 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#pragma once
+
+// clang-format off
+
+#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
+#define CPU_CAPABILITY_AVX512
+#endif
+
+#include <ATen/cpu/vec/functional.h>
+#include <ATen/cpu/vec/vec.h>
+
+namespace {
+
+using namespace at::vec;
+
+template <typename scalar_t,
+          typename std::enable_if_t<is_reduced_floating_point_v<scalar_t>, int> = 0>
+inline Vectorized<scalar_t> convert_from_float_ext(const Vectorized<float>& a, const Vectorized<float>& b) {
+  return at::vec::convert_from_float<scalar_t>(a, b);
+}
+
+#if defined(CPU_CAPABILITY_AVX512)
+
+// `at::vec::convert_from_float<>` from PyTorch doesn't have avx512-bf16 intrinsics
+// use native instruction for bfloat16->float32 conversion
+template <>
+inline Vectorized<at::BFloat16> convert_from_float_ext<at::BFloat16>(const Vectorized<float>& a, const Vectorized<float>& b) {
+  return (__m512i)(_mm512_cvtne2ps_pbh(__m512(b), __m512(a)));
+}
+
+#define CVT_BF16_TO_FP32(a) \
+    _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(a), 16))
+
+#define CVT_FP16_TO_FP32(a) \
+    _mm512_cvtps_ph(a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))
+
+// this doesn't hanel NaN.
+inline __m512bh cvt_e4m3_bf16_intrinsic_no_nan(__m256i fp8_vec) {
+  const __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
+
+  const __m512i mant = _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(0x07)), 4);
+  const __m512i raw_exp = _mm512_srli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(0x78)), 3);
+  const __m512i exp = _mm512_slli_epi16(_mm512_add_epi16(raw_exp, _mm512_set1_epi16(120)), 7);
+  const __m512i nonsign = _mm512_or_si512(exp, mant);
+
+  const __m512i sign = _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(0x80)), 8);
+  const __m512i combined = _mm512_or_si512(nonsign, sign);
+
+  const __mmask32 is_nonzero = _mm512_cmpneq_epi16_mask(x, _mm512_setzero_si512());
+  return (__m512bh)_mm512_maskz_mov_epi16(is_nonzero, combined);
+}
+
+inline __m512bh cvt_e4m3_bf16_intrinsic_without_denorm(__m256i fp8_vec) {
+  // The following conversion is without denorm behavior, that is to say,
+  //   Max subnorm   : S.0000.111 = 0.875 ∗ 2**(−6)
+  //   Min subnorm   : S.0000.001 = 2**(−9)
+  // 0.0019 ~ 0.0137 cannot be converted correctly.
+  __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
+  auto mask = _mm512_cmpneq_epi16_mask(
+      _mm512_and_si512(x, _mm512_set1_epi16(127)),
+      _mm512_setzero_si512());  // mask = x & 0x7f
+  auto mask_nan = _mm512_cmpneq_epi16_mask(
+      _mm512_and_si512(x, _mm512_set1_epi16(127)),
+      _mm512_set1_epi16(127));                                                      // mask_nan = x & 0x7f
+  auto mantissa = _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(7)), 4);  // mantissa = (x & 7) << 4
+  auto exponent = _mm512_add_epi16(
+      _mm512_srli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(120)), 3),
+      _mm512_set1_epi16(120));  // exponent = (((x >> 3) & 15) + 120)
+  auto nonsign = _mm512_maskz_mov_epi16(mask, _mm512_or_si512(mantissa, _mm512_slli_epi16(exponent, 7)));
+  nonsign = _mm512_mask_mov_epi16(_mm512_set1_epi16(0x7fff), mask_nan, nonsign);  // deal with Nan
+  return (__m512bh)(_mm512_or_si512(
+      nonsign,
+      _mm512_slli_epi16(
+          _mm512_and_si512(x, _mm512_set1_epi16(128)),
+          8)));  // add sign (x & 128) << 8
+}
+
+inline __m512bh cvt_e4m3_bf16_intrinsic_with_denorm(__m256i fp8_vec) {
+  __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
+  __m512i lg2mant = _mm512_mask_mov_epi16(
+      _mm512_mask_mov_epi16(
+          _mm512_setzero_si512(), _mm512_test_epi16_mask(x, _mm512_set1_epi16(2)), _mm512_set1_epi16(1)),
+      _mm512_test_epi16_mask(x, _mm512_set1_epi16(4)),
+      _mm512_set1_epi16(2));
+  return (__m512bh)(_mm512_or_si512(
+      _mm512_maskz_mov_epi16(
+          _mm512_cmpneq_epi16_mask(_mm512_and_si512(x, _mm512_set1_epi16(127)), _mm512_setzero_si512()),
+          _mm512_mask_blend_epi16(
+              _mm512_test_epi16_mask(x, _mm512_set1_epi16(120)),
+              _mm512_or_si512(
+                  _mm512_and_si512(
+                      _mm512_sllv_epi16(
+                          _mm512_and_si512(x, _mm512_set1_epi16(3)), _mm512_sub_epi16(_mm512_set1_epi16(7), lg2mant)),
+                      _mm512_set1_epi16(0x007f)),
+                  _mm512_slli_epi16(_mm512_add_epi16(lg2mant, _mm512_set1_epi16(118)), 7)),
+              _mm512_or_si512(
+                  _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(7)), 4),
+                  _mm512_slli_epi16(
+                      _mm512_add_epi16(
+                          _mm512_srli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(120)), 3), _mm512_set1_epi16(120)),
+                      7)))),
+      _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(128)), 8)));
+}
+
+inline __m512bh CVT_FP8_TO_BF16(__m256i a) {
+#ifdef SGLANG_CPU_FP8_CVT_FTZ
+  return cvt_e4m3_bf16_intrinsic_no_nan(a);
+#else
+  return cvt_e4m3_bf16_intrinsic_with_denorm(a);
+#endif
+}
+
+#endif
+
+// vector to scalar reduction
+#if defined(CPU_CAPABILITY_AVX512) && 0
+inline float vec_reduce_sum(const Vectorized<float>& a) {
+  return _mm512_reduce_add_ps(__m512(a));
+}
+
+inline float vec_reduce_max(const Vectorized<float>& a) {
+  return _mm512_reduce_max_ps(__m512(a));
+}
+#else
+inline float vec_reduce_sum(const Vectorized<float>& a) {
+  return vec_reduce_all([](Vectorized<float>& x, Vectorized<float>& y) { return x + y; }, a);
+}
+
+inline float vec_reduce_max(const Vectorized<float>& a) {
+  return vec_reduce_all([](Vectorized<float>& x, Vectorized<float>& y) { return maximum(x, y); }, a);
+}
+#endif
+
+// https://github.com/InternLM/lmdeploy/blob/086481ed84b59bee3b8e4274e5fc69620040c048/lmdeploy/pytorch/kernels/cuda/w8a8_triton_kernels.py#L282
+template <typename scalar_t>
+inline void quantize_row_int8(uint8_t* __restrict__ Aq, float& As,
+    const scalar_t* __restrict__ A, int64_t K, float eps = 1e-7) {
+
+  float amax = 0.f; // absolute max
+  for (int64_t k = 0; k < K; ++k) {
+    const float val = static_cast<float>(A[k]);
+    amax = std::max(amax, std::abs(val));
+  }
+
+  amax = std::max(amax, eps);
+  const float scale = amax / 127;
+  const float inv_scale = 127 / amax;
+
+  for (int64_t k = 0; k < K; ++k) {
+    const float val = static_cast<float>(A[k]) * inv_scale;
+    Aq[k] = (uint8_t)(std::round(val)) + 128;
+  }
+  As = scale;
+}
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <>
+inline void quantize_row_int8<at::BFloat16>(uint8_t* __restrict__ Aq, float& As,
+    const at::BFloat16* __restrict__ A, int64_t K, float eps) {
+
+  const __m512 signBit = _mm512_set1_ps(-0.0f);
+  const __m512i off = _mm512_set1_epi32(128);
+
+  // K is 32x, no remainder
+  float amax = 0.f;
+  __m512 vamax0 = _mm512_set1_ps(0.f);
+  __m512 vamax1 = _mm512_set1_ps(0.f);
+  for (int64_t k = 0; k < K; k += 32) {
+    __m512i va = _mm512_loadu_si512((void*)(A + k));
+    __m512 va0 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 0));
+    __m512 va1 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 1));
+    vamax0 = _mm512_max_ps(vamax0, _mm512_andnot_ps(signBit, va0));
+    vamax1 = _mm512_max_ps(vamax1, _mm512_andnot_ps(signBit, va1));
+  }
+  amax = _mm512_reduce_max_ps(_mm512_max_ps(vamax0, vamax1));
+  amax = std::max(amax, eps);
+  const float scale = amax / 127;
+  const float inv_scale = 127 / amax;
+  const __m512 vd = _mm512_set1_ps(inv_scale);
+
+  for (int64_t k = 0; k < K; k += 32) {
+    __m512i va = _mm512_loadu_si512((void*)(A + k));
+    __m512 va0 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 0));
+    __m512 va1 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 1));
+    va0 = _mm512_mul_ps(va0, vd);
+    va1 = _mm512_mul_ps(va1, vd);
+    va0 = _mm512_roundscale_ps(va0, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));
+    va1 = _mm512_roundscale_ps(va1, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));
+    __m128i i0 = _mm512_cvtepi32_epi8(_mm512_add_epi32(_mm512_cvtps_epi32(va0), off));
+    __m128i i1 = _mm512_cvtepi32_epi8(_mm512_add_epi32(_mm512_cvtps_epi32(va1), off));
+    _mm256_storeu_si256(reinterpret_cast<__m256i*>(Aq + k), _mm256_set_m128i(i1, i0));
+  }
+  As = scale;
+}
+#endif
+
+// transpose utils
+// taken from my PR in ggml: https://github.com/ggml-org/llama.cpp/pull/8998
+#if defined(CPU_CAPABILITY_AVX512)
+inline void transpose_16x16_32bit(__m512i * v) {
+  __m512i v1[16];
+  v1[0] = _mm512_unpacklo_epi32(v[0], v[1]);
+  v1[1] = _mm512_unpackhi_epi32(v[0], v[1]);
+  v1[2] = _mm512_unpacklo_epi32(v[2], v[3]);
+  v1[3] = _mm512_unpackhi_epi32(v[2], v[3]);
+  v1[4] = _mm512_unpacklo_epi32(v[4], v[5]);
+  v1[5] = _mm512_unpackhi_epi32(v[4], v[5]);
+  v1[6] = _mm512_unpacklo_epi32(v[6], v[7]);
+  v1[7] = _mm512_unpackhi_epi32(v[6], v[7]);
+  v1[8] = _mm512_unpacklo_epi32(v[8], v[9]);
+  v1[9] = _mm512_unpackhi_epi32(v[8], v[9]);
+  v1[10] = _mm512_unpacklo_epi32(v[10], v[11]);
+  v1[11] = _mm512_unpackhi_epi32(v[10], v[11]);
+  v1[12] = _mm512_unpacklo_epi32(v[12], v[13]);
+  v1[13] = _mm512_unpackhi_epi32(v[12], v[13]);
+  v1[14] = _mm512_unpacklo_epi32(v[14], v[15]);
+  v1[15] = _mm512_unpackhi_epi32(v[14], v[15]);
+
+  v[0] = _mm512_unpacklo_epi64(v1[0], v1[2]);
+  v[1] = _mm512_unpackhi_epi64(v1[0], v1[2]);
+  v[2] = _mm512_unpacklo_epi64(v1[1], v1[3]);
+  v[3] = _mm512_unpackhi_epi64(v1[1], v1[3]);
+  v[4] = _mm512_unpacklo_epi64(v1[4], v1[6]);
+  v[5] = _mm512_unpackhi_epi64(v1[4], v1[6]);
+  v[6] = _mm512_unpacklo_epi64(v1[5], v1[7]);
+  v[7] = _mm512_unpackhi_epi64(v1[5], v1[7]);
+  v[8] = _mm512_unpacklo_epi64(v1[8], v1[10]);
+  v[9] = _mm512_unpackhi_epi64(v1[8], v1[10]);
+  v[10] = _mm512_unpacklo_epi64(v1[9], v1[11]);
+  v[11] = _mm512_unpackhi_epi64(v1[9], v1[11]);
+  v[12] = _mm512_unpacklo_epi64(v1[12], v1[14]);
+  v[13] = _mm512_unpackhi_epi64(v1[12], v1[14]);
+  v[14] = _mm512_unpacklo_epi64(v1[13], v1[15]);
+  v[15] = _mm512_unpackhi_epi64(v1[13], v1[15]);
+
+  v1[0] = _mm512_shuffle_i32x4(v[0], v[4], 0x88);
+  v1[1] = _mm512_shuffle_i32x4(v[1], v[5], 0x88);
+  v1[2] = _mm512_shuffle_i32x4(v[2], v[6], 0x88);
+  v1[3] = _mm512_shuffle_i32x4(v[3], v[7], 0x88);
+  v1[4] = _mm512_shuffle_i32x4(v[0], v[4], 0xdd);
+  v1[5] = _mm512_shuffle_i32x4(v[1], v[5], 0xdd);
+  v1[6] = _mm512_shuffle_i32x4(v[2], v[6], 0xdd);
+  v1[7] = _mm512_shuffle_i32x4(v[3], v[7], 0xdd);
+  v1[8] = _mm512_shuffle_i32x4(v[8], v[12], 0x88);
+  v1[9] = _mm512_shuffle_i32x4(v[9], v[13], 0x88);
+  v1[10] = _mm512_shuffle_i32x4(v[10], v[14], 0x88);
+  v1[11] = _mm512_shuffle_i32x4(v[11], v[15], 0x88);
+  v1[12] = _mm512_shuffle_i32x4(v[8], v[12], 0xdd);
+  v1[13] = _mm512_shuffle_i32x4(v[9], v[13], 0xdd);
+  v1[14] = _mm512_shuffle_i32x4(v[10], v[14], 0xdd);
+  v1[15] = _mm512_shuffle_i32x4(v[11], v[15], 0xdd);
+
+  v[0] = _mm512_shuffle_i32x4(v1[0], v1[8], 0x88);
+  v[1] = _mm512_shuffle_i32x4(v1[1], v1[9], 0x88);
+  v[2] = _mm512_shuffle_i32x4(v1[2], v1[10], 0x88);
+  v[3] = _mm512_shuffle_i32x4(v1[3], v1[11], 0x88);
+  v[4] = _mm512_shuffle_i32x4(v1[4], v1[12], 0x88);
+  v[5] = _mm512_shuffle_i32x4(v1[5], v1[13], 0x88);
+  v[6] = _mm512_shuffle_i32x4(v1[6], v1[14], 0x88);
+  v[7] = _mm512_shuffle_i32x4(v1[7], v1[15], 0x88);
+  v[8] = _mm512_shuffle_i32x4(v1[0], v1[8], 0xdd);
+  v[9] = _mm512_shuffle_i32x4(v1[1], v1[9], 0xdd);
+  v[10] = _mm512_shuffle_i32x4(v1[2], v1[10], 0xdd);
+  v[11] = _mm512_shuffle_i32x4(v1[3], v1[11], 0xdd);
+  v[12] = _mm512_shuffle_i32x4(v1[4], v1[12], 0xdd);
+  v[13] = _mm512_shuffle_i32x4(v1[5], v1[13], 0xdd);
+  v[14] = _mm512_shuffle_i32x4(v1[6], v1[14], 0xdd);
+  v[15] = _mm512_shuffle_i32x4(v1[7], v1[15], 0xdd);
+}
+
+// remove warning : ignoring attributes on template argument ‘__m512i’ [-Wignored-attributes]
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wignored-attributes"
+
+// transpose from [2, 32] to [32, 2]
+inline std::tuple<__m512i, __m512i> transpose_2x32_16bit(__m512i r0, __m512i r1) {
+  // r0: {a0, a1, ..., a31}
+  // r1: {b0, b1, ..., b31}
+  //
+  // d0: {a0,   b0, ..., a15, b15}
+  // d1: {a16, b16, ..., a31, b31}
+  //
+  __m512i d0 = _mm512_unpacklo_epi16(r0, r1);
+  __m512i d1 = _mm512_unpackhi_epi16(r0, r1);
+  r0 = _mm512_shuffle_i32x4(d0, d1, 0x88);
+  r1 = _mm512_shuffle_i32x4(d0, d1, 0xdd);
+  d0 = _mm512_shuffle_i32x4(r0, r1, 0x88);
+  d1 = _mm512_shuffle_i32x4(r0, r1, 0xdd);
+  return std::make_tuple(d0, d1);
+}
+#pragma GCC diagnostic pop
+
+#endif
+
+// TODO: debug print, remove me later
+template<typename scalar_t>
+void print_array(scalar_t* ptr, int size) {
+  for (int d = 0; d < size; ++d) {
+    if (d % 16 == 0) { std::cout << std::endl; }
+    std::cout << ptr[d] << " ";
+  }
+  std::cout << std::endl;
+}
+
+} // anonymous namespace

csrc/cpu/shm.cpp

@@ -7,9 +7,10 @@
 
 namespace {
 #define MAX_SHM_RANK_NUM 8
-#define MAX_THREAD_NUM 12
-#define PER_THREAD_SHM_BUFFER_BYTES (4 * 1024 * 1024)
-#define MIN_THREAD_PROCESS_SIZE (8 * 1024)
+#define PER_THREAD_SHM_BUFFER_BYTES (2 * 1024 * 1024)
+static_assert(PER_THREAD_SHM_BUFFER_BYTES % 2 == 0);
+#define PER_THREAD_SHM_BUFFER_OFFSET (PER_THREAD_SHM_BUFFER_BYTES >> 1)
+#define MIN_THREAD_PROCESS_SIZE (256)
 #define MAX_P2P_SEND_TENSOR_NUM 8
 
 template <typename scalar_t>
@@ -32,10 +33,10 @@ struct KernelVecType<c10::Half> {
   using scalar_vec_t = vec_op::FP16Vec16;
 };
 
-enum class ThreadSHMStat : char { THREAD_READY = 0, SHM_DATA_READY, DONE };
-
 struct ThreadSHMContext {
-  volatile ThreadSHMStat thread_stats[MAX_SHM_RANK_NUM];
+  volatile char _curr_thread_stamp;
+  volatile char _ready_thread_stamp;
+  char _padding1[6];
   int thread_id;
   int thread_num;
   int rank;
@@ -44,14 +45,19 @@ struct ThreadSHMContext {
   int swizzled_ranks[MAX_SHM_RANK_NUM];
   void* thread_shm_ptrs[MAX_SHM_RANK_NUM];
   ThreadSHMContext* shm_contexts[MAX_SHM_RANK_NUM];
+  size_t _thread_buffer_mask;
+  char _padding2[56];
 
   ThreadSHMContext(const int thread_id, const int thread_num, const int rank,
                    const int group_size, void* thread_shm_ptr)
-      : thread_id(thread_id),
+      : _curr_thread_stamp(1),
+        _ready_thread_stamp(0),
+        thread_id(thread_id),
         thread_num(thread_num),
         rank(rank),
         group_size(group_size),
-        _spinning_count(0) {
+        _spinning_count(0),
+        _thread_buffer_mask(0) {
     static_assert(sizeof(ThreadSHMContext) % 64 == 0);
     TORCH_CHECK(group_size <= MAX_SHM_RANK_NUM);
     TORCH_CHECK((size_t)this % 64 == 0);
@@ -60,7 +66,6 @@ struct ThreadSHMContext {
       shm_contexts[i] = nullptr;
       thread_shm_ptrs[i] = nullptr;
       swizzled_ranks[i] = (i + rank) % group_size;
-      thread_stats[i] = ThreadSHMStat::DONE;
     }
     set_context(rank, this, thread_shm_ptr);
   }
@@ -77,59 +82,66 @@ struct ThreadSHMContext {
 
   template <typename T>
   T* get_thread_shm_ptr(int rank) {
-    return reinterpret_cast<T*>(thread_shm_ptrs[rank]);
+    return reinterpret_cast<T*>(
+        reinterpret_cast<int8_t*>(thread_shm_ptrs[rank]) +
+        (PER_THREAD_SHM_BUFFER_OFFSET & _thread_buffer_mask));
+  }
+
+  void next_buffer() { _thread_buffer_mask ^= 0xFFFFFFFFFFFFFFFF; }
+
+  char get_curr_stamp() const { return _curr_thread_stamp; }
+
+  char get_ready_stamp() const { return _ready_thread_stamp; }
+
+  void next_stamp() {
+    _mm_mfence();
+    _curr_thread_stamp += 1;
+  }
+
+  void commit_ready_stamp() {
+    _mm_mfence();
+    _ready_thread_stamp = _curr_thread_stamp;
   }
 
   int get_swizzled_rank(int idx) { return swizzled_ranks[idx]; }
 
-  void wait_for_all(ThreadSHMStat prev_stat) {
-    for (int idx = 0; idx < group_size; ++idx) {
+  template <typename Cond>
+  void wait_for_all(Cond&& cond) {
+    for (int idx = 1; idx < group_size; ++idx) {
       int rank = get_swizzled_rank(idx);
-      while (thread_stats[rank] == prev_stat) {
-        ++_spinning_count;
-        _mm_pause();
-      }
+      wait_for_one(rank, std::forward<Cond>(cond));
     }
-    vec_op::mem_barrier();
   }
 
-  void wait_for_one(int rank, ThreadSHMStat prev_stat) {
-    while (thread_stats[rank] == prev_stat) {
+  template <typename Cond>
+  void wait_for_one(int rank, Cond&& cond) {
+    ThreadSHMContext* rank_ctx = shm_contexts[rank];
+    for (;;) {
+      char local_curr_stamp = get_curr_stamp();
+      char local_ready_stamp = get_ready_stamp();
+      char rank_curr_stamp = rank_ctx->get_curr_stamp();
+      char rank_ready_stamp = rank_ctx->get_ready_stamp();
+      if (cond(local_curr_stamp, local_ready_stamp, rank_curr_stamp,
+               rank_ready_stamp)) {
+        break;
+      }
       ++_spinning_count;
       _mm_pause();
     }
-    vec_op::mem_barrier();
   }
 
-  void set_thread_stat(ThreadSHMStat stat) {
-    for (int idx = 0; idx < group_size; ++idx) {
-      int rank = get_swizzled_rank(idx);
-      shm_contexts[rank]->thread_stats[this->rank] = stat;
-    }
+  static bool check_no_buffer_conflict(char local_curr_stamp,
+                                       char local_ready_stamp,
+                                       char rank_curr_stamp,
+                                       char rank_ready_stamp) {
+    char temp = rank_curr_stamp + 2;
+    return local_curr_stamp != temp;
   }
 
-  void set_thread_stat(int target_rank, ThreadSHMStat stat) {
-    for (int idx = 0; idx < group_size; ++idx) {
-      int rank = get_swizzled_rank(idx);
-      shm_contexts[rank]->thread_stats[target_rank] = stat;
-    }
-  }
-
-  // barrier for all ranks in the group, used for all2all ops
-  // DONE -> THREAD_READY -> SHM_DATA_READY -> DONE -> ...
-  void barrier(ThreadSHMStat next_stat) {
-    if (next_stat == ThreadSHMStat::THREAD_READY) {
-      set_thread_stat(ThreadSHMStat::THREAD_READY);
-      wait_for_all(ThreadSHMStat::DONE);
-    } else if (next_stat == ThreadSHMStat::SHM_DATA_READY) {
-      set_thread_stat(ThreadSHMStat::SHM_DATA_READY);
-      wait_for_all(ThreadSHMStat::THREAD_READY);
-    } else if (next_stat == ThreadSHMStat::DONE) {
-      set_thread_stat(ThreadSHMStat::DONE);
-      wait_for_all(ThreadSHMStat::SHM_DATA_READY);
-    } else {
-      TORCH_CHECK(false, "Invalid next_stat to barrier.");
-    }
+  static bool check_stamp_ready(char local_curr_stamp, char local_ready_stamp,
+                                char rank_curr_stamp, char rank_ready_stamp) {
+    char temp = local_curr_stamp + 1;
+    return (local_curr_stamp == rank_ready_stamp) || (temp == rank_ready_stamp);
   }
 
   std::string to_string() const {
@@ -164,7 +176,7 @@ class SHMManager {
                       const int group_size)
       : _rank(rank),
         _group_size(group_size),
-        _thread_num(std::min(torch::get_num_threads(), MAX_THREAD_NUM)),
+        _thread_num(torch::get_num_threads()),
         _shm_names({""}),
         _shared_mem_ptrs({nullptr}),
         _shm_ctx(nullptr) {
@@ -326,7 +338,8 @@ void shm_cc_loop(ThreadSHMContext* ctx, int64_t elem_num, F&& inner_func) {
       (total_units_num + thread_num - 1) / thread_num;
   int64_t per_unit_elem_num = MIN_THREAD_PROCESS_SIZE / sizeof(scalar_t);
   int64_t max_per_thread_iteration_elem_num =
-      PER_THREAD_SHM_BUFFER_BYTES / sizeof(scalar_t);
+      (PER_THREAD_SHM_BUFFER_BYTES >> 1) /
+      sizeof(scalar_t);  // Note: double buffer
   int64_t per_thread_elem_num = per_unit_elem_num * per_thread_units_num;
 
 #pragma omp parallel for schedule(static, 1)
@@ -336,10 +349,13 @@ void shm_cc_loop(ThreadSHMContext* ctx, int64_t elem_num, F&& inner_func) {
     int64_t curr_elem_num =
         std::min(max_per_thread_iteration_elem_num, end - offset);
     ThreadSHMContext* thread_ctx = ctx + i;
+    bool fast_mode = ((end - offset) <= max_per_thread_iteration_elem_num);
 
     while (curr_elem_num > 0) {
-      inner_func(thread_ctx, offset, curr_elem_num);
+      inner_func(thread_ctx, offset, curr_elem_num, fast_mode);
 
+      thread_ctx->next_stamp();
+      thread_ctx->next_buffer();
       offset += max_per_thread_iteration_elem_num;
       curr_elem_num = std::min(max_per_thread_iteration_elem_num, end - offset);
     }
@@ -397,7 +413,7 @@ void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
   shm_cc_ops::shm_cc_loop<scalar_t>(
       ctx, elem_num,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
+          int64_t data_elem_num, bool fast_mode) {
         int rank = thread_ctx->rank;
         scalar_t* thread_shm_ptr =
             thread_ctx->get_thread_shm_ptr<scalar_t>(rank);
@@ -410,16 +426,17 @@ void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
               thread_ctx->get_swizzled_rank(idx + 1));
         });
 
-        thread_ctx->barrier(ThreadSHMStat::THREAD_READY);
+        if (!fast_mode) {
+          thread_ctx->wait_for_all(ThreadSHMContext::check_no_buffer_conflict);
+        }
 
         shm_cc_ops::memcpy_to_shm(thread_shm_ptr, thread_data_ptr,
                                   thread_data_elem_num);
-
-        thread_ctx->barrier(ThreadSHMStat::SHM_DATA_READY);
-
+        thread_ctx->commit_ready_stamp();
         int64_t aligned_data_elem_num =
             (data_elem_num / vec_elem_num) * vec_elem_num;
         int64_t i = 0;
+        thread_ctx->wait_for_all(ThreadSHMContext::check_stamp_ready);
 #pragma GCC unroll 4
         for (; i < aligned_data_elem_num; i += vec_elem_num) {
           vec_t local_data(thread_data_ptr + i);  // load from cache
@@ -447,8 +464,6 @@ void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
           reduced_data.save(thread_data_ptr + i,
                             data_elem_num - aligned_data_elem_num);
         }
-
-        thread_ctx->barrier(ThreadSHMStat::DONE);
       });
 
   return;
@@ -488,18 +503,18 @@ void shm_gather_impl(ThreadSHMContext* ctx, scalar_t* data, size_t elem_num,
   shm_cc_ops::shm_cc_loop<scalar_t>(
       ctx, elem_num,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
+          int64_t data_elem_num, bool fast_mode) {
         int rank = thread_ctx->rank;
         scalar_t* thread_shm_ptr =
             thread_ctx->get_thread_shm_ptr<scalar_t>(rank);
 
-        thread_ctx->barrier(ThreadSHMStat::THREAD_READY);
-
-        shm_cc_ops::memcpy_to_shm(thread_shm_ptr, data + data_offset,
-                                  data_elem_num * sizeof(scalar_t));
-
-        thread_ctx->barrier(ThreadSHMStat::SHM_DATA_READY);
+        if (!fast_mode) {
+          thread_ctx->wait_for_all(ThreadSHMContext::check_no_buffer_conflict);
+        }
 
+        shm_cc_ops::memcpy(thread_shm_ptr, data + data_offset,
+                           data_elem_num * sizeof(scalar_t));
+        thread_ctx->commit_ready_stamp();
         if (rank == dst) {
           shm_cc_ops::memcpy(outputs[rank] + data_offset, data + data_offset,
                              data_elem_num * sizeof(scalar_t));
@@ -508,12 +523,12 @@ void shm_gather_impl(ThreadSHMContext* ctx, scalar_t* data, size_t elem_num,
             scalar_t* src_ptr =
                 thread_ctx->get_thread_shm_ptr<scalar_t>(src_rank);  // shm
             scalar_t* dst_ptr = outputs[src_rank] + data_offset;
-            shm_cc_ops::memcpy_from_shm(dst_ptr, src_ptr,
-                                        data_elem_num * sizeof(scalar_t));
+            thread_ctx->wait_for_one(src_rank,
+                                     ThreadSHMContext::check_stamp_ready);
+            shm_cc_ops::memcpy(dst_ptr, src_ptr,
+                               data_elem_num * sizeof(scalar_t));
           }
         }
-
-        thread_ctx->barrier(ThreadSHMStat::DONE);
       });
 
   return;
@@ -599,7 +614,7 @@ struct TensorListMeta {
   int8_t _padding[40];
 };
 
-void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
+void shm_send_tensor_list_impl(ThreadSHMContext* ctx, int64_t dst,
                                const std::vector<torch::Tensor>& tensor_list) {
   CPU_KERNEL_GUARD_IN(shm_send_tensor_list_impl)
   std::vector<torch::Tensor> tensor_list_with_metadata;
@@ -620,12 +635,11 @@ void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
   shm_cc_ops::shm_cc_loop<int8_t>(
       ctx, metadata->total_bytes,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
+          int64_t data_elem_num, bool fast_mode) {
         int rank = thread_ctx->rank;
-        // Wait until the receiver set the stat to DONE
-        thread_ctx->wait_for_one(rank, ThreadSHMStat::SHM_DATA_READY);
-
         int64_t curr_shm_offset = 0;
+        thread_ctx->wait_for_one(dst,
+                                 ThreadSHMContext::check_no_buffer_conflict);
         while (curr_shm_offset < data_elem_num) {
           MemPiece frag = metadata->get_data(data_offset + curr_shm_offset);
           frag.size = std::min(frag.size, data_elem_num - curr_shm_offset);
@@ -634,8 +648,7 @@ void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
               frag.ptr, frag.size);
           curr_shm_offset += frag.size;
         }
-
-        thread_ctx->set_thread_stat(rank, ThreadSHMStat::SHM_DATA_READY);
+        thread_ctx->commit_ready_stamp();
       });
 }
 
@@ -646,8 +659,7 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
   torch::Tensor metadata_tensor =
       torch::empty({sizeof(TensorListMeta)}, options);
 
-  // Wait until the sender set the stat of the thread 0 to SHM_DATA_READY
-  ctx->wait_for_one(src, ThreadSHMStat::DONE);
+  ctx->wait_for_one(src, ThreadSHMContext::check_stamp_ready);
   shm_cc_ops::memcpy(metadata_tensor.data_ptr(),
                      ctx->get_thread_shm_ptr<void>(src),
                      sizeof(TensorListMeta));
@@ -664,9 +676,8 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
   shm_cc_ops::shm_cc_loop<int8_t>(
       ctx, metadata.total_bytes,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
-        // Wait until the sender set the stat to SHM_DATA_READY
-        thread_ctx->wait_for_one(src, ThreadSHMStat::DONE);
+          int64_t data_elem_num, bool fast_mode) {
+        ctx->wait_for_one(src, ThreadSHMContext::check_stamp_ready);
         int64_t curr_shm_offset = 0;
         while (curr_shm_offset < data_elem_num) {
           MemPiece frag = metadata.get_data(data_offset + curr_shm_offset);
@@ -677,8 +688,6 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
               frag.size);
           curr_shm_offset += frag.size;
         }
-
-        thread_ctx->set_thread_stat(src, ThreadSHMStat::DONE);
       });
 
   std::vector<torch::Tensor> tensor_list;
@@ -756,7 +765,8 @@ void shm_send_tensor_list(int64_t handle,
                           int64_t dst) {
   CPU_KERNEL_GUARD_IN(shm_send_tensor_list)
   shm_send_tensor_list_impl(
-      SHMManager::get_singleton_instance(handle)->get_shm_ctx(), tensor_list);
+      SHMManager::get_singleton_instance(handle)->get_shm_ctx(), dst,
+      tensor_list);
   CPU_KERNEL_GUARD_OUT(shm_send_tensor_list)
 }
 
@@ -778,4 +788,4 @@ std::string join_shm_manager(int64_t handle, const std::string& name) {
   TORCH_CHECK(shm_manager);
   shm_manager->join(name);
   return shm_manager->get_shm_ctx()->to_string();
-}
+}

csrc/cpu/torch_bindings.cpp

@@ -50,6 +50,27 @@ void shm_send_tensor_list(int64_t handle,
 
 std::vector<torch::Tensor> shm_recv_tensor_list(int64_t handle, int64_t src);
 
+at::Tensor weight_packed_linear(at::Tensor& mat1, at::Tensor& mat2,
+                                const std::optional<at::Tensor>& bias,
+                                bool is_vnni);
+
+at::Tensor convert_weight_packed(at::Tensor& weight);
+
+at::Tensor fused_experts_cpu(
+    at::Tensor& hidden_states, at::Tensor& w1, at::Tensor& w2,
+    at::Tensor& topk_weights, at::Tensor& topk_ids, bool inplace,
+    bool use_int8_w8a8, bool use_fp8_w8a16,
+    const std::optional<at::Tensor>& w1_scale,
+    const std::optional<at::Tensor>& w2_scale,
+    const std::optional<std::vector<int64_t>> block_size,
+    const std::optional<at::Tensor>& a1_scale,
+    const std::optional<at::Tensor>& a2_scale, bool is_vnni);
+
+at::Tensor int8_scaled_mm_with_quant(at::Tensor& mat1, at::Tensor& mat2,
+                                     at::Tensor& scales2,
+                                     const std::optional<at::Tensor>& bias,
+                                     at::ScalarType out_dtype, bool is_vnni);
+
 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // vLLM custom ops
 
@@ -214,6 +235,28 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("shm_recv_tensor_list(int handle, int src) -> Tensor[](a)",
           &shm_recv_tensor_list);
 #endif
+
+  // sgl-kernels
+#if defined(__AVX512BF16__) && defined(__AVX512F__) && defined(__AVX512VNNI__)
+  ops.def(
+      "weight_packed_linear(Tensor(a0!) mat1, Tensor(a1!) mat2, Tensor(a2!)? "
+      "bias, bool is_vnni) -> Tensor");
+  ops.impl("weight_packed_linear", torch::kCPU, &weight_packed_linear);
+  ops.def("convert_weight_packed(Tensor! weight) -> Tensor");
+  ops.impl("convert_weight_packed", torch::kCPU, &convert_weight_packed);
+  ops.def(
+      "fused_experts_cpu(Tensor! hidden_states, Tensor w1, Tensor w2, Tensor "
+      "topk_weights, Tensor topk_ids, bool inplace, bool use_int8_w8a8, bool "
+      "use_fp8_w8a16, Tensor? w1_scale, Tensor? w2_scale, SymInt[]? "
+      "block_size, Tensor? a1_scale, Tensor? a2_scale, bool is_vnni) -> "
+      "Tensor");
+  ops.impl("fused_experts_cpu", torch::kCPU, &fused_experts_cpu);
+  ops.def(
+      "int8_scaled_mm_with_quant(Tensor mat1, Tensor mat2, Tensor scales2, "
+      "Tensor? bias, ScalarType out_dtype, bool is_vnni) -> Tensor");
+  ops.impl("int8_scaled_mm_with_quant", torch::kCPU,
+           &int8_scaled_mm_with_quant);
+#endif
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {

csrc/moe/marlin_moe_wna16/marlin_template.h

@@ -1255,8 +1255,6 @@ __global__ void Marlin(
     if constexpr (has_zp && !is_zp_float) {
       if (is_new_zp) {
         if constexpr (group_blocks == -1) is_first_matmul_in_slice = false;
-        FragB frag_zp_0;
-        FragB frag_zp_1;
         int zp_quant_0, zp_quant_1;
 
         if constexpr (w_type.size_bits() == 4) {

csrc/moe/moe_align_sum_kernels.cu

@@ -239,7 +239,7 @@ void moe_sum(torch::Tensor& input,   // [num_tokens, topk, hidden_size]
              torch::Tensor& output)  // [num_tokens, hidden_size]
 {
   const int hidden_size = input.size(-1);
-  const int num_tokens = output.numel() / hidden_size;
+  const auto num_tokens = output.numel() / hidden_size;
   const int topk = input.size(1);
 
   dim3 grid(num_tokens);

csrc/moe/topk_softmax_kernels.cu

@@ -492,7 +492,7 @@ void topk_softmax(
     torch::Tensor& gating_output)               // [num_tokens, num_experts]
 {
     const int num_experts = gating_output.size(-1);
-    const int num_tokens = gating_output.numel() / num_experts;
+    const auto num_tokens = gating_output.numel() / num_experts;
     const int topk = topk_weights.size(-1);
 
     const bool is_pow_2 = (num_experts != 0) && ((num_experts & (num_experts - 1)) == 0);

csrc/quantization/cutlass_w8a8/c3x/scaled_mm.cuh

@@ -144,4 +144,65 @@ struct cutlass_3x_gemm_sm100 {
       Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>;
 };
 
+template <typename ElementAB_, typename ElementD_,
+          template <typename, typename, typename> typename Epilogue_,
+          typename TileShape, typename ClusterShape, typename KernelSchedule,
+          typename EpilogueSchedule>
+struct cutlass_3x_gemm_sm120 {
+  using ElementAB = ElementAB_;
+  using LayoutA = cutlass::layout::RowMajor;
+  static constexpr int AlignmentA =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+
+  using LayoutB = cutlass::layout::ColumnMajor;
+  static constexpr int AlignmentB =
+      128 / cutlass::sizeof_bits<ElementAB>::value;
+
+  using ElementC = void;
+  using LayoutC = cutlass::layout::RowMajor;
+  static constexpr int AlignmentC =
+      128 / cutlass::sizeof_bits<ElementD_>::value;
+
+  using ElementD = ElementD_;
+  using LayoutD = cutlass::layout::RowMajor;
+  static constexpr int AlignmentD = AlignmentC;
+
+  using ElementAcc =
+      typename std::conditional<std::is_same_v<ElementAB, int8_t>, int32_t,
+                                float>::type;
+  using Epilogue = Epilogue_<ElementAcc, ElementD, TileShape>;
+
+  // MMA type
+  using ElementAccumulator = float;
+
+  // Epilogue types
+  using ElementBias = cutlass::half_t;
+  using ElementCompute = float;
+  using ElementAux = ElementD;
+  using LayoutAux = LayoutD;
+  using ElementAmax = float;
+
+  using EVTCompute = typename Epilogue::EVTCompute;
+
+  using CollectiveEpilogue =
+      typename cutlass::epilogue::collective::CollectiveBuilder<
+          cutlass::arch::Sm120, cutlass::arch::OpClassTensorOp, TileShape,
+          ClusterShape, cutlass::epilogue::collective::EpilogueTileAuto,
+          ElementAccumulator, ElementCompute, ElementC, LayoutC, AlignmentC,
+          ElementD, LayoutD, AlignmentD, EpilogueSchedule,
+          EVTCompute>::CollectiveOp;
+
+  using CollectiveMainloop =
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          cutlass::arch::Sm120, cutlass::arch::OpClassTensorOp, ElementAB,
+          LayoutA, AlignmentA, ElementAB, LayoutB, AlignmentB,
+          ElementAccumulator, TileShape, ClusterShape,
+          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
+              sizeof(typename CollectiveEpilogue::SharedStorage))>,
+          KernelSchedule>::CollectiveOp;
+
+  using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
+      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue, void>;
+};
+
 }  // namespace vllm

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_kernels.hpp

@@ -36,6 +36,12 @@ void cutlass_scaled_mm_sm100_fp8(torch::Tensor& out, torch::Tensor const& a,
                                  torch::Tensor const& b_scales,
                                  std::optional<torch::Tensor> const& bias);
 
+void cutlass_scaled_mm_sm120_fp8(torch::Tensor& out, torch::Tensor const& a,
+                                 torch::Tensor const& b,
+                                 torch::Tensor const& a_scales,
+                                 torch::Tensor const& b_scales,
+                                 std::optional<torch::Tensor> const& bias);
+
 void cutlass_scaled_mm_blockwise_sm100_fp8(torch::Tensor& out,
                                            torch::Tensor const& a,
                                            torch::Tensor const& b,

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu

@@ -0,0 +1,24 @@
+#include "scaled_mm_kernels.hpp"
+#include "scaled_mm_sm120_fp8_dispatch.cuh"
+#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
+
+namespace vllm {
+
+void cutlass_scaled_mm_sm120_fp8(torch::Tensor& out, torch::Tensor const& a,
+                                 torch::Tensor const& b,
+                                 torch::Tensor const& a_scales,
+                                 torch::Tensor const& b_scales,
+                                 std::optional<torch::Tensor> const& bias) {
+  TORCH_CHECK(a_scales.is_contiguous() && b_scales.is_contiguous());
+  if (bias) {
+    TORCH_CHECK(bias->dtype() == out.dtype(),
+                "currently bias dtype must match output dtype ", out.dtype());
+    return cutlass_scaled_mm_sm120_fp8_epilogue<c3x::ScaledEpilogueBias>(
+        out, a, b, a_scales, b_scales, *bias);
+  } else {
+    return cutlass_scaled_mm_sm120_fp8_epilogue<c3x::ScaledEpilogue>(
+        out, a, b, a_scales, b_scales);
+  }
+}
+
+}  // namespace vllm

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8_dispatch.cuh

@@ -0,0 +1,67 @@
+#pragma once
+
+#include "scaled_mm.cuh"
+#include "cutlass_gemm_caller.cuh"
+
+/**
+ * This file defines Gemm kernel configurations for SM120 (fp8) based on the
+ * Gemm shape.
+ */
+
+namespace vllm {
+
+using c3x::cutlass_gemm_caller;
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue>
+struct sm120_fp8_config_default {
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  using KernelSchedule = cutlass::gemm::collective::KernelScheduleAuto;
+  using EpilogueSchedule = cutlass::epilogue::collective::EpilogueScheduleAuto;
+  using TileShape = Shape<_128, _128, _128>;
+  using ClusterShape = Shape<_1, _1, _1>;  // Only work with Shape<_1, _1, _1>
+  using Cutlass3xGemm =
+      cutlass_3x_gemm_sm120<InType, OutType, Epilogue, TileShape, ClusterShape,
+                            KernelSchedule, EpilogueSchedule>;
+};
+
+template <typename InType, typename OutType,
+          template <typename, typename, typename> typename Epilogue,
+          typename... EpilogueArgs>
+inline void cutlass_gemm_sm120_fp8_dispatch(torch::Tensor& out,
+                                            torch::Tensor const& a,
+                                            torch::Tensor const& b,
+                                            EpilogueArgs&&... args) {
+  static_assert(std::is_same<InType, cutlass::float_e4m3_t>());
+  TORCH_CHECK(a.dtype() == torch::kFloat8_e4m3fn);
+  TORCH_CHECK(b.dtype() == torch::kFloat8_e4m3fn);
+
+  using Cutlass3xGemmDefault =
+      typename sm120_fp8_config_default<InType, OutType,
+                                        Epilogue>::Cutlass3xGemm;
+  return cutlass_gemm_caller<Cutlass3xGemmDefault>(
+      out, a, b, std::forward<EpilogueArgs>(args)...);
+}
+
+template <template <typename, typename, typename> typename Epilogue,
+          typename... EpilogueArgs>
+void cutlass_scaled_mm_sm120_fp8_epilogue(torch::Tensor& out,
+                                          torch::Tensor const& a,
+                                          torch::Tensor const& b,
+                                          EpilogueArgs&&... epilogue_args) {
+  TORCH_CHECK(a.dtype() == torch::kFloat8_e4m3fn);
+  TORCH_CHECK(b.dtype() == torch::kFloat8_e4m3fn);
+
+  if (out.dtype() == torch::kBFloat16) {
+    return cutlass_gemm_sm120_fp8_dispatch<cutlass::float_e4m3_t,
+                                           cutlass::bfloat16_t, Epilogue>(
+        out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
+  } else {
+    TORCH_CHECK(out.dtype() == torch::kFloat16);
+    return cutlass_gemm_sm120_fp8_dispatch<cutlass::float_e4m3_t,
+                                           cutlass::half_t, Epilogue>(
+        out, a, b, std::forward<EpilogueArgs>(epilogue_args)...);
+  }
+}
+
+}  // namespace vllm

csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu

@@ -0,0 +1,34 @@
+#include <cudaTypedefs.h>
+#include "c3x/scaled_mm_kernels.hpp"
+
+#include "cuda_utils.h"
+
+/*
+   This file defines quantized GEMM operations using the CUTLASS 3.x API, for
+   NVIDIA GPUs with sm120 (Blackwell Geforce).
+*/
+
+#if defined ENABLE_SCALED_MM_SM120 && ENABLE_SCALED_MM_SM120
+
+void cutlass_scaled_mm_sm120(torch::Tensor& c, torch::Tensor const& a,
+                             torch::Tensor const& b,
+                             torch::Tensor const& a_scales,
+                             torch::Tensor const& b_scales,
+                             std::optional<torch::Tensor> const& bias) {
+  TORCH_CHECK(a_scales.dtype() == torch::kFloat32);
+  TORCH_CHECK(b_scales.dtype() == torch::kFloat32);
+
+  int M = a.size(0), N = b.size(1), K = a.size(1);
+  TORCH_CHECK(
+      (a_scales.numel() == 1 || a_scales.numel() == a.size(0)) &&
+          (b_scales.numel() == 1 || b_scales.numel() == b.size(1)),
+      "Currently, block scaled fp8 gemm is not implemented for Blackwell");
+
+  // Standard per-tensor/per-token/per-channel scaling
+  TORCH_CHECK(a_scales.is_contiguous() && b_scales.is_contiguous());
+  TORCH_CHECK(a.dtype() == torch::kFloat8_e4m3fn,
+              "Currently, only fp8 gemm is implemented for Blackwell");
+  vllm::cutlass_scaled_mm_sm120_fp8(c, a, b, a_scales, b_scales, bias);
+}
+
+#endif

csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu

@@ -41,6 +41,14 @@ void cutlass_moe_mm_sm90(
 
 #endif
 
+#if defined ENABLE_SCALED_MM_SM120 && ENABLE_SCALED_MM_SM120
+void cutlass_scaled_mm_sm120(torch::Tensor& c, torch::Tensor const& a,
+                             torch::Tensor const& b,
+                             torch::Tensor const& a_scales,
+                             torch::Tensor const& b_scales,
+                             std::optional<torch::Tensor> const& bias);
+#endif
+
 #if defined ENABLE_SCALED_MM_SM100 && ENABLE_SCALED_MM_SM100
 void cutlass_scaled_mm_sm100(torch::Tensor& c, torch::Tensor const& a,
                              torch::Tensor const& b,
@@ -168,8 +176,15 @@ void cutlass_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
   at::cuda::OptionalCUDAGuard const device_guard(device_of(a));
   int32_t version_num = get_sm_version_num();
 
+#if defined ENABLE_SCALED_MM_SM120 && ENABLE_SCALED_MM_SM120
+  if (version_num >= 120) {
+    cutlass_scaled_mm_sm120(c, a, b, a_scales, b_scales, bias);
+    return;
+  }
+#endif
+
 #if defined ENABLE_SCALED_MM_SM100 && ENABLE_SCALED_MM_SM100
-  if (version_num >= 100) {
+  if (version_num >= 100 && version_num < 120) {
     cutlass_scaled_mm_sm100(c, a, b, a_scales, b_scales, bias);
     return;
   }

csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu

@@ -335,8 +335,10 @@ void run_fp4_blockwise_scaled_group_mm(
   TORCH_CHECK(status == cutlass::Status::kSuccess, "Failed to run GEMM");
 }
 
+#if defined ENABLE_NVFP4 && ENABLE_NVFP4
 constexpr auto FLOAT4_E2M1X2 = at::ScalarType::Byte;
 constexpr auto SF_DTYPE = at::ScalarType::Float8_e4m3fn;
+#endif
 
 #define CHECK_TYPE(x, st, m) \
   TORCH_CHECK(x.scalar_type() == st, ": Inconsistency of Tensor type:", m)

csrc/quantization/gptq_marlin/marlin_template.h

@@ -1113,8 +1113,6 @@ __global__ void Marlin(
     if constexpr (has_zp && !is_zp_float) {
       if (is_new_zp) {
         if constexpr (group_blocks == -1) is_first_matmul_in_slice = false;
-        FragB frag_zp_0;
-        FragB frag_zp_1;
         int zp_quant_0, zp_quant_1;
 
         if constexpr (w_type.size_bits() == 4) {

docs/.nav.yml

@@ -39,6 +39,7 @@ nav:
       - models/generative_models.md
       - models/pooling_models.md
       - models/extensions
+      - Hardware Supported Models: models/hardware_supported_models
     - Features:
       - features/compatibility_matrix.md
       - features/*

docs/README.md

@@ -1,7 +1,8 @@
 # Welcome to vLLM
 
 <figure markdown="span">
-  ![](./assets/logos/vllm-logo-text-light.png){ align="center" alt="vLLM" class="no-scaled-link" width="60%" }
+  ![](./assets/logos/vllm-logo-text-light.png){ align="center" alt="vLLM Light" class="logo-light" width="60%" }
+  ![](./assets/logos/vllm-logo-text-dark.png){ align="center" alt="vLLM Dark" class="logo-dark" width="60%" }
 </figure>
 
 <p style="text-align:center">

docs/configuration/engine_args.md

@@ -6,7 +6,7 @@ title: Engine Arguments
 Engine arguments control the behavior of the vLLM engine.
 
 - For [offline inference][offline-inference], they are part of the arguments to [LLM][vllm.LLM] class.
-- For [online serving][openai-compatible-server], they are part of the arguments to `vllm serve`.
+- For [online serving][serving-openai-compatible-server], they are part of the arguments to `vllm serve`.
 
 You can look at [EngineArgs][vllm.engine.arg_utils.EngineArgs] and [AsyncEngineArgs][vllm.engine.arg_utils.AsyncEngineArgs] to see the available engine arguments.
 

docs/contributing/deprecation_policy.md

@@ -37,14 +37,14 @@ multiple Y releases:
 - **Timeline**: A removal version is explicitly stated in the deprecation
 warning (e.g., "This will be removed in v0.10.0").
 - **Communication**: Deprecation is noted in the following, as applicable:
-  - Help strings
-  - Log output
-  - API responses
-  - `/metrics` output (for metrics features)
-  - User-facing documentation
-  - Release notes
-  - GitHub Issue (RFC) for feedback
-  - Documentation and use of the `@typing_extensions.deprecated` decorator for Python APIs
+    - Help strings
+    - Log output
+    - API responses
+    - `/metrics` output (for metrics features)
+    - User-facing documentation
+    - Release notes
+    - GitHub Issue (RFC) for feedback
+    - Documentation and use of the `@typing_extensions.deprecated` decorator for Python APIs
 
 **2.Deprecated (Off By Default)**
 

docs/contributing/model/multimodal.md

@@ -538,11 +538,13 @@ return a schema of the tensors outputted by the HF processor that are related to
             prompt: str,
             mm_data: Mapping[str, object],
             mm_kwargs: Mapping[str, object],
+            tok_kwargs: Mapping[str, object],
         ) -> BatchFeature:
             processed_outputs = super()._call_hf_processor(
                 prompt=prompt,
                 mm_data=mm_data,
                 mm_kwargs=mm_kwargs,
+                tok_kwargs=tok_kwargs,
             )
 
             image_patches = processed_outputs.get("image_patches")
@@ -566,6 +568,11 @@ return a schema of the tensors outputted by the HF processor that are related to
         Our [actual code](gh-file:vllm/model_executor/models/fuyu.py) has special handling
         for text-only inputs to prevent unnecessary warnings from HF processor.
 
+    !!! note
+        The `_call_hf_processor` method specifies both `mm_kwargs` and `tok_kwargs` for
+        processing. `mm_kwargs` is used to both initialize and call the huggingface
+        processor, whereas `tok_kwargs` is only used to call the huggingface processor.
+
     This lets us override [_get_mm_fields_config][vllm.multimodal.processing.BaseMultiModalProcessor._get_mm_fields_config] as follows:
 
     ```python

docs/design/arch_overview.md

@@ -74,7 +74,7 @@ python -m vllm.entrypoints.openai.api_server --model <model>
 
 That code can be found in <gh-file:vllm/entrypoints/openai/api_server.py>.
 
-More details on the API server can be found in the [OpenAI-Compatible Server][openai-compatible-server] document.
+More details on the API server can be found in the [OpenAI-Compatible Server][serving-openai-compatible-server] document.
 
 ## LLM Engine
 

docs/design/v1/prefix_caching.md

@@ -117,8 +117,8 @@ There are two design points to highlight:
 
 1. We allocate all KVCacheBlock when initializing the KV cache manager to be a block pool. This avoids Python object creation overheads and can easily track all blocks all the time.  
 2. We introduce doubly linked list pointers directly in the KVCacheBlock, so that we could construct a free queue directly. This gives us two benefits:  
-   1. We could have O(1) complexity moving elements in the middle to the tail.  
-   2. We could avoid introducing another Python queue (e.g., `deque`) which has a wrapper to the elements.
+    1. We could have O(1) complexity moving elements in the middle to the tail.  
+    2. We could avoid introducing another Python queue (e.g., `deque`) which has a wrapper to the elements.
 
 As a result, we will have the following components when the KV cache manager is initialized:
 
@@ -135,19 +135,19 @@ As a result, we will have the following components when the KV cache manager is
 
 **New request:** Workflow for the scheduler to schedule a new request with KV cache block allocation:
 
-1. The scheduler calls `kv_cache_manager.get_computed_blocks()` to get a sequence of blocks that have already been computed. This is done by hashing the prompt tokens in the request and looking up Cache Blocks.  
+1. The scheduler calls `kv_cache_manager.get_computed_blocks()` to get a sequence of blocks that have already been computed. This is done by hashing the prompt tokens in the request and looking up cache blocks.  
 2. The scheduler calls `kv_cache_manager.allocate_slots()`. It does the following steps:  
-   1. Compute the number of new required blocks, and return if there are no sufficient blocks to allocate.  
-   2. “Touch” the computed blocks. It increases the reference count of the computed block by one, and removes the block from the free queue if the block wasn’t used by other requests. This is to avoid these computed blocks being evicted. See the example in the next section for illustration.  
-   3. Allocate new blocks by popping the heads of the free queue. If the head block is a cached block, this also “evicts” the block so that no other requests can reuse it anymore from now on.  
-   4. If an allocated block is already full of tokens, we immediately add it to the Cache Block, so that the block can be reused by other requests in the same batch.
+    1. Compute the number of new required blocks, and return if there are no sufficient blocks to allocate.  
+    2. “Touch” the computed blocks. It increases the reference count of the computed block by one, and removes the block from the free queue if the block wasn’t used by other requests. This is to avoid these computed blocks being evicted. See the example in the next section for illustration.  
+    3. Allocate new blocks by popping the heads of the free queue. If the head block is a cached block, this also “evicts” the block so that no other requests can reuse it anymore from now on.  
+    4. If an allocated block is already full of tokens, we immediately add it to the cache block, so that the block can be reused by other requests in the same batch.
 
 **Running request:** Workflow for the scheduler to schedule a running request with KV cache block allocation:
 
 1. The scheduler calls `kv_cache_manager.allocate_slots()`. It does the following steps:  
-   1. Compute the number of new required blocks, and return if there are no sufficient blocks to allocate.  
-   2. Allocate new blocks by popping the heads of the free queue. If the head block is a cached block, this also “evicts” the block so that no other requests can reuse it anymore from now on.  
-   3. Append token IDs to the slots in existing blocks as well as the new blocks. If a block is full, we add it to the Cache Block to cache it.
+    1. Compute the number of new required blocks, and return if there are no sufficient blocks to allocate.  
+    2. Allocate new blocks by popping the heads of the free queue. If the head block is a cached block, this also “evicts” the block so that no other requests can reuse it anymore from now on.  
+    3. Append token IDs to the slots in existing blocks as well as the new blocks. If a block is full, we add it to the cache block to cache it.
 
 **Duplicated blocks**  
 Assuming block size is 4 and you send a request (Request 1\) with prompt ABCDEF and decoding length 3:
@@ -199,7 +199,7 @@ When a request is finished, we free all its blocks if no other requests are usin
 When the head block (least recently used block) of the free queue is cached, we have to evict the block to prevent it from being used by other requests. Specifically, eviction involves the following steps:
 
 1. Pop the block from the head of the free queue. This is the LRU block to be evicted.  
-2. Remove the block ID from the Cache Block.  
+2. Remove the block ID from the cache block.  
 3. Remove the block hash.
 
 ## Example

docs/features/compatibility_matrix.md

@@ -59,23 +59,23 @@ th:not(:first-child) {
 
 ## Feature x Hardware
 
-| Feature                                                   | Volta              | Turing   | Ampere   | Ada   | Hopper   | CPU                | AMD   |
-|-----------------------------------------------------------|--------------------|----------|----------|-------|----------|--------------------|-------|
-| [CP][chunked-prefill]                                     | [❌](gh-issue:2729) | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     |
-| [APC][automatic-prefix-caching]                           | [❌](gh-issue:3687) | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     |
-| [LoRA][lora-adapter]                                      | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     |
-| <abbr title="Prompt Adapter">prmpt adptr</abbr>           | ✅                  | ✅        | ✅        | ✅     | ✅        | [❌](gh-issue:8475) | ✅     |
-| [SD][spec-decode]                                         | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     |
-| CUDA graph                                                | ✅                  | ✅        | ✅        | ✅     | ✅        | ❌                  | ✅     |
-| <abbr title="Pooling Models">pooling</abbr>               | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ❔     |
-| <abbr title="Encoder-Decoder Models">enc-dec</abbr>       | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ❌     |
-| <abbr title="Multimodal Inputs">mm</abbr>                 | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     |
-| <abbr title="Logprobs">logP</abbr>                        | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     |
-| <abbr title="Prompt Logprobs">prmpt logP</abbr>           | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     |
-| <abbr title="Async Output Processing">async output</abbr> | ✅                  | ✅        | ✅        | ✅     | ✅        | ❌                  | ❌     |
-| multi-step                                                | ✅                  | ✅        | ✅        | ✅     | ✅        | [❌](gh-issue:8477) | ✅     |
-| best-of                                                   | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     |
-| beam-search                                               | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     |
+| Feature                                                   | Volta               | Turing    | Ampere    | Ada    | Hopper     | CPU                | AMD    | TPU |
+|-----------------------------------------------------------|---------------------|-----------|-----------|--------|------------|--------------------|--------|-----|
+| [CP][chunked-prefill]                                     | [❌](gh-issue:2729) | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅ |
+| [APC][automatic-prefix-caching]                           | [❌](gh-issue:3687) | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅ |
+| [LoRA][lora-adapter]                                      | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ✅ |
+| <abbr title="Prompt Adapter">prmpt adptr</abbr>           | ✅                  | ✅        | ✅        | ✅     | ✅        | [❌](gh-issue:8475) | ✅     | ❌ |
+| [SD][spec-decode]                                         | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ❌ |
+| CUDA graph                                                | ✅                  | ✅        | ✅        | ✅     | ✅        | ❌                  | ✅     | ❌ |
+| <abbr title="Pooling Models">pooling</abbr>               | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ❔     | ❌ |
+| <abbr title="Encoder-Decoder Models">enc-dec</abbr>       | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ❌     | ❌ |
+| <abbr title="Multimodal Inputs">mm</abbr>                 | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ❌ |
+| <abbr title="Logprobs">logP</abbr>                        | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ❌ |
+| <abbr title="Prompt Logprobs">prmpt logP</abbr>           | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ❌ |
+| <abbr title="Async Output Processing">async output</abbr> | ✅                  | ✅        | ✅        | ✅     | ✅        | ❌                  | ❌     | ❌ |
+| multi-step                                                | ✅                  | ✅        | ✅        | ✅     | ✅        | [❌](gh-issue:8477) | ✅     | ❌ |
+| best-of                                                   | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ❌ |
+| beam-search                                               | ✅                  | ✅        | ✅        | ✅     | ✅        | ✅                  | ✅     | ❌ |
 
 !!! note
     Please refer to [Feature support through NxD Inference backend][feature-support-through-nxd-inference-backend] for features supported on AWS Neuron hardware

docs/features/structured_outputs.md

@@ -21,7 +21,7 @@ The following parameters are supported, which must be added as extra parameters:
 - `guided_grammar`: the output will follow the context free grammar.
 - `structural_tag`: Follow a JSON schema within a set of specified tags within the generated text.
 
-You can see the complete list of supported parameters on the [OpenAI-Compatible Server][openai-compatible-server] page.
+You can see the complete list of supported parameters on the [OpenAI-Compatible Server][serving-openai-compatible-server] page.
 
 Structured outputs are supported by default in the OpenAI-Compatible Server. You
 may choose to specify the backend to use by setting the

docs/features/tool_calling.md

@@ -53,7 +53,7 @@ Next, make a request to the model that should result in it using the available t
     tool_call = response.choices[0].message.tool_calls[0].function
     print(f"Function called: {tool_call.name}")
     print(f"Arguments: {tool_call.arguments}")
-    print(f"Result: {get_weather(**json.loads(tool_call.arguments))}")
+    print(f"Result: {tool_functions[tool_call.name](**json.loads(tool_call.arguments))}")
     ```
 
 Example output:
@@ -99,6 +99,14 @@ vLLM supports the `tool_choice='required'` option in the chat completion API. Si
 
 When tool_choice='required' is set, the model is guaranteed to generate one or more tool calls based on the specified tool list in the `tools` parameter. The number of tool calls depends on the user's query. The output format strictly follows the schema defined in the `tools` parameter.
 
+## None Function Calling
+
+vLLM supports the `tool_choice='none'` option in the chat completion API. When this option is set, the model will not generate any tool calls and will respond with regular text content only, even if tools are defined in the request.
+
+By default, when `tool_choice='none'` is specified, vLLM excludes tool definitions from the prompt to optimize context usage. To include tool definitions even with `tool_choice='none'`, use the `--expand-tools-even-if-tool-choice-none` option.
+
+Note: This behavior will change in v0.10.0, where tool definitions will be included by default even with `tool_choice='none'`.
+
 ## Automatic Function Calling
 
 To enable this feature, you should set the following flags:

docs/getting_started/installation/cpu.md

@@ -118,6 +118,7 @@ vLLM CPU backend supports the following vLLM features:
 - `VLLM_CPU_OMP_THREADS_BIND`: specify the CPU cores dedicated to the OpenMP threads. For example, `VLLM_CPU_OMP_THREADS_BIND=0-31` means there will be 32 OpenMP threads bound on 0-31 CPU cores. `VLLM_CPU_OMP_THREADS_BIND=0-31|32-63` means there will be 2 tensor parallel processes, 32 OpenMP threads of rank0 are bound on 0-31 CPU cores, and the OpenMP threads of rank1 are bound on 32-63 CPU cores. By setting to `auto`, the OpenMP threads of each rank are bound to the CPU cores in each NUMA node. By setting to `all`, the OpenMP threads of each rank uses all CPU cores available on the system. Default value is `auto`.
 - `VLLM_CPU_NUM_OF_RESERVED_CPU`: specify the number of CPU cores which are not dedicated to the OpenMP threads for each rank. The variable only takes effect when VLLM_CPU_OMP_THREADS_BIND is set to `auto`. Default value is `0`.
 - `VLLM_CPU_MOE_PREPACK`: whether to use prepack for MoE layer. This will be passed to `ipex.llm.modules.GatedMLPMOE`. Default is `1` (True). On unsupported CPUs, you might need to set this to `0` (False).
+- `VLLM_CPU_SGL_KERNEL` (Experimental): whether to use small-batch optimized kernels for linear layer and MoE layer, especially for low-latency requirements like online serving. The kernels require AMX instruction set, BFloat16 weight type and weight shapes divisible by 32. Default is `0` (False).
 
 ## Performance tips
 

docs/getting_started/installation/google_tpu.md

@@ -58,9 +58,9 @@ assigned to your Google Cloud project for your immediate exclusive use.
 
 For more information about using TPUs with GKE, see:
 
-- <https://cloud.google.com/kubernetes-engine/docs/how-to/tpus>
-- <https://cloud.google.com/kubernetes-engine/docs/concepts/tpus>
-- <https://cloud.google.com/kubernetes-engine/docs/concepts/plan-tpus>
+- [About TPUs in GKE](https://cloud.google.com/kubernetes-engine/docs/concepts/tpus)
+- [Deploy TPU workloads in GKE Standard](https://cloud.google.com/kubernetes-engine/docs/how-to/tpus)
+- [Plan for TPUs in GKE](https://cloud.google.com/kubernetes-engine/docs/concepts/plan-tpus)
 
 ## Configure a new environment
 

docs/getting_started/installation/intel_gaudi.md

@@ -110,7 +110,7 @@ docker run \
 ### Supported features
 
 - [Offline inference][offline-inference]
-- Online serving via [OpenAI-Compatible Server][openai-compatible-server]
+- Online serving via [OpenAI-Compatible Server][serving-openai-compatible-server]
 - HPU autodetection - no need to manually select device within vLLM
 - Paged KV cache with algorithms enabled for Intel Gaudi accelerators
 - Custom Intel Gaudi implementations of Paged Attention, KV cache ops,

docs/mkdocs/javascript/slack_and_forum.js

@@ -0,0 +1,56 @@
+/**
+ * slack_and_forum.js
+ *
+ * Adds a custom Slack and Forum button to the MkDocs Material header.
+ *
+ */
+
+window.addEventListener('DOMContentLoaded', () => {
+  const headerInner = document.querySelector('.md-header__inner');
+
+  if (headerInner) {
+    const slackButton = document.createElement('button');
+    slackButton.className = 'slack-button';
+    slackButton.title = 'Join us on Slack';
+    slackButton.style.border = 'none';
+    slackButton.style.background = 'transparent';
+    slackButton.style.cursor = 'pointer';
+
+    slackButton.innerHTML = `
+      <img src="https://a.slack-edge.com/80588/marketing/img/icons/icon_slack_hash_colored.png" 
+           style="height: 1.1rem;" 
+           alt="Slack">
+    `;
+
+    slackButton.addEventListener('click', () => {
+      window.open('https://slack.vllm.ai', '_blank', 'noopener');
+    });
+
+    const forumButton = document.createElement('button');
+    forumButton.className = 'forum-button';
+    forumButton.title = 'Join the Forum';
+    forumButton.style.border = 'none';
+    forumButton.style.background = 'transparent';
+    forumButton.style.cursor = 'pointer';
+
+    forumButton.innerHTML = `
+      <svg
+        xmlns="http://www.w3.org/2000/svg"
+        viewBox="0 -960 960 960"
+        fill="currentColor"
+      >
+        <path d="M817.85-198.15 698.46-317.54H320q-24.48 0-41.47-16.99T261.54-376v-11.69h424.61q25.39 0 43.47-18.08 18.07-18.08 18.07-43.46v-268.92h11.69q24.48 0 41.47 16.99 17 16.99 17 41.47v461.54ZM179.08-434.69l66.84-66.85h363.31q10.77 0 17.69-6.92 6.93-6.92 6.93-17.69v-246.77q0-10.77-6.93-17.7-6.92-6.92-17.69-6.92H203.69q-10.77 0-17.69 6.92-6.92 6.93-6.92 17.7v338.23Zm-36.93 89.46v-427.69q0-25.39 18.08-43.46 18.08-18.08 43.46-18.08h405.54q25.39 0 43.46 18.08 18.08 18.07 18.08 43.46v246.77q0 25.38-18.08 43.46-18.07 18.07-43.46 18.07H261.54L142.15-345.23Zm36.93-180.92V-797.54v271.39Z"/>
+      </svg>
+    `;
+
+    forumButton.addEventListener('click', () => {
+      window.open('https://discuss.vllm.ai/', '_blank', 'noopener');
+    });
+
+    const githubSource = document.querySelector('.md-header__source');
+    if (githubSource) {
+      githubSource.parentNode.insertBefore(slackButton, githubSource.nextSibling);
+      githubSource.parentNode.insertBefore(forumButton, slackButton.nextSibling);
+    }
+  }
+});

docs/mkdocs/stylesheets/extra.css

@@ -108,3 +108,38 @@ body[data-md-color-scheme="slate"] .md-nav__item--section > label.md-nav__link .
 .md-content__button-wrapper a:hover {
   color: var(--md-accent-fg-color);
 }
+
+/* Slack and Forum css */
+.slack-button, 
+.forum-button {
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  margin-left: 0.4rem;
+  height: 24px;
+}
+
+.slack-button img {
+  height: 18px;
+  filter: none !important;
+}
+
+.slack-button:hover,
+.forum-button:hover {
+  opacity: 0.7;
+}
+
+.forum-button svg {
+  height: 28px;
+  opacity: 0.9;
+  transform: translateY(2px);
+}
+
+/* For logo css */
+[data-md-color-scheme="default"] .logo-dark {
+  display: none;
+}
+
+[data-md-color-scheme="slate"] .logo-light {
+  display: none;
+}

docs/models/generative_models.md

@@ -134,7 +134,7 @@ outputs = llm.chat(conversation, chat_template=custom_template)
 
 ## Online Serving
 
-Our [OpenAI-Compatible Server][openai-compatible-server] provides endpoints that correspond to the offline APIs:
+Our [OpenAI-Compatible Server][serving-openai-compatible-server] provides endpoints that correspond to the offline APIs:
 
 - [Completions API][completions-api] is similar to `LLM.generate` but only accepts text.
 - [Chat API][chat-api]  is similar to `LLM.chat`, accepting both text and [multi-modal inputs][multimodal-inputs] for models with a chat template.

docs/models/hardware_supported_models/tpu.md

@@ -0,0 +1,36 @@
+---
+title: TPU
+---
+[](){ #tpu-supported-models }
+
+# TPU Supported Models
+## Text-only Language Models
+
+| Model                                               | Architecture                   | Supported |
+|-----------------------------------------------------|--------------------------------|-----------|
+| mistralai/Mixtral-8x7B-Instruct-v0.1                | MixtralForCausalLM             | 🟨 |
+| mistralai/Mistral-Small-24B-Instruct-2501           | MistralForCausalLM             | ✅ |
+| mistralai/Codestral-22B-v0.1                        | MistralForCausalLM             | ✅ |
+| mistralai/Mixtral-8x22B-Instruct-v0.1               | MixtralForCausalLM             | ❌ |
+| meta-llama/Llama-3.3-70B-Instruct                   | LlamaForCausalLM               | ✅ |
+| meta-llama/Llama-3.1-8B-Instruct                    | LlamaForCausalLM               | ✅ |
+| meta-llama/Llama-3.1-70B-Instruct                   | LlamaForCausalLM               | ✅ |
+| meta-llama/Llama-4-*                                | Llama4ForConditionalGeneration | ❌ |
+| microsoft/Phi-3-mini-128k-instruct                  | Phi3ForCausalLM                | 🟨 |
+| microsoft/phi-4                                     | Phi3ForCausalLM                | ❌ |
+| google/gemma-3-27b-it                               | Gemma3ForConditionalGeneration | 🟨 |
+| google/gemma-3-4b-it                                | Gemma3ForConditionalGeneration | ❌ |
+| deepseek-ai/DeepSeek-R1                             | DeepseekV3ForCausalLM          | ❌ |
+| deepseek-ai/DeepSeek-V3                             | DeepseekV3ForCausalLM          | ❌ |
+| RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8  | LlamaForCausalLM               | ✅ |
+| RedHatAI/Meta-Llama-3.1-70B-Instruct-quantized.w8a8 | LlamaForCausalLM               | ✅ |
+| Qwen/Qwen3-8B                                       | Qwen3ForCausalLM               | ✅ |
+| Qwen/Qwen3-32B                                      | Qwen3ForCausalLM               | ✅ |
+| Qwen/Qwen2.5-7B-Instruct                            | Qwen2ForCausalLM               | ✅ |
+| Qwen/Qwen2.5-32B                                    | Qwen2ForCausalLM               | ✅ |
+| Qwen/Qwen2.5-14B-Instruct                           | Qwen2ForCausalLM               | ✅ |
+| Qwen/Qwen2.5-1.5B-Instruct                          | Qwen2ForCausalLM               | 🟨 |
+
+✅ Runs and optimized.  
+🟨 Runs and correct but not optimized to green yet.  
+❌ Does not pass accuracy test or does not run.  

docs/models/pooling_models.md

@@ -113,7 +113,7 @@ A code example can be found here: <gh-file:examples/offline_inference/basic/scor
 
 ## Online Serving
 
-Our [OpenAI-Compatible Server][openai-compatible-server] provides endpoints that correspond to the offline APIs:
+Our [OpenAI-Compatible Server][serving-openai-compatible-server] provides endpoints that correspond to the offline APIs:
 
 - [Pooling API][pooling-api] is similar to `LLM.encode`, being applicable to all types of pooling models.
 - [Embeddings API][embeddings-api] is similar to `LLM.embed`, accepting both text and [multi-modal inputs][multimodal-inputs] for embedding models.

docs/models/supported_models.md

@@ -34,7 +34,7 @@ llm.apply_model(lambda model: print(type(model)))
 If it is `TransformersForCausalLM` then it means it's based on Transformers!
 
 !!! tip
-    You can force the use of `TransformersForCausalLM` by setting `model_impl="transformers"` for [offline-inference][offline-inference] or `--model-impl transformers` for the [openai-compatible-server][openai-compatible-server].
+    You can force the use of `TransformersForCausalLM` by setting `model_impl="transformers"` for [offline-inference][offline-inference] or `--model-impl transformers` for the [openai-compatible-server][serving-openai-compatible-server].
 
 !!! note
     vLLM may not fully optimise the Transformers implementation so you may see degraded performance if comparing a native model to a Transformers model in vLLM.
@@ -53,8 +53,8 @@ For a model to be compatible with the Transformers backend for vLLM it must:
 
 If the compatible model is:
 
-- on the Hugging Face Model Hub, simply set `trust_remote_code=True` for [offline-inference][offline-inference] or `--trust-remote-code` for the [openai-compatible-server][openai-compatible-server].
-- in a local directory, simply pass directory path to `model=<MODEL_DIR>` for [offline-inference][offline-inference] or `vllm serve <MODEL_DIR>` for the [openai-compatible-server][openai-compatible-server].
+- on the Hugging Face Model Hub, simply set `trust_remote_code=True` for [offline-inference][offline-inference] or `--trust-remote-code` for the [openai-compatible-server][serving-openai-compatible-server].
+- in a local directory, simply pass directory path to `model=<MODEL_DIR>` for [offline-inference][offline-inference] or `vllm serve <MODEL_DIR>` for the [openai-compatible-server][serving-openai-compatible-server].
 
 This means that, with the Transformers backend for vLLM, new models can be used before they are officially supported in Transformers or vLLM!
 
@@ -329,6 +329,9 @@ Specified using `--task generate`.
 | `DeepseekForCausalLM`                             | DeepSeek                                            | `deepseek-ai/deepseek-llm-67b-base`, `deepseek-ai/deepseek-llm-7b-chat` etc.                                                                                                 |                        | ✅︎                          | ✅︎                     |
 | `DeepseekV2ForCausalLM`                           | DeepSeek-V2                                         | `deepseek-ai/DeepSeek-V2`, `deepseek-ai/DeepSeek-V2-Chat` etc.                                                                                                               |                        | ✅︎                          | ✅︎                     |
 | `DeepseekV3ForCausalLM`                           | DeepSeek-V3                                         | `deepseek-ai/DeepSeek-V3-Base`, `deepseek-ai/DeepSeek-V3` etc.                                                                                                               |                        | ✅︎                          | ✅︎                     |
+| `Dots1ForCausalLM`                                | dots.llm1                                           | `rednote-hilab/dots.llm1.base`, `rednote-hilab/dots.llm1.inst` etc.                                                                                                               |                        | ✅︎                          | ✅︎                     |
+| `Ernie4_5_ForCausalLM`                            | Ernie4.5                                            | `baidu/ERNIE-4.5-0.3B-PT`,etc.                                                                                                               |                        | ✅︎                          | ✅︎                     |
+| `Ernie4_5_MoeForCausalLM`                         | Ernie4.5MoE                                         | `baidu/ERNIE-4.5-21B-A3B-PT`,  `baidu/ERNIE-4.5-300B-A47B-PT`, etc.                                                                                                               |                        | ✅︎                          | ✅︎                     |
 | `ExaoneForCausalLM`                               | EXAONE-3                                            | `LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct`, etc.                                                                                                                                 | ✅︎                     | ✅︎                          | ✅︎                     |
 | `FalconForCausalLM`                               | Falcon                                              | `tiiuae/falcon-7b`, `tiiuae/falcon-40b`, `tiiuae/falcon-rw-7b`, etc.                                                                                                         |                        | ✅︎                          | ✅︎                     |
 | `FalconMambaForCausalLM`                          | FalconMamba                                         | `tiiuae/falcon-mamba-7b`, `tiiuae/falcon-mamba-7b-instruct`, etc.                                                                                                            |                        | ✅︎                          | ✅︎                     |
@@ -349,6 +352,7 @@ Specified using `--task generate`.
 | `GraniteMoeSharedForCausalLM`                     | Granite MoE Shared                                  | `ibm-research/moe-7b-1b-active-shared-experts` (test model)                                                                                                                  | ✅︎                     | ✅︎                          | ✅︎                     |
 | `GritLM`                                          | GritLM                                              | `parasail-ai/GritLM-7B-vllm`.                                                                                                                                                | ✅︎                     | ✅︎                          |                       |
 | `Grok1ModelForCausalLM`                           | Grok1                                               | `hpcai-tech/grok-1`.                                                                                                                                                         | ✅︎                     | ✅︎                          | ✅︎                     |
+| `HunYuanMoEV1ForCausalLM`                         | Hunyuan-80B-A13B                                    | `tencent/Hunyuan-A13B-Instruct`,  `tencent/Hunyuan-A13B-Pretrain`, `tencent/Hunyuan-A13B-Instruct-FP8`etc.                                                                  |                        |                             | ✅︎                     |
 | `InternLMForCausalLM`                             | InternLM                                            | `internlm/internlm-7b`, `internlm/internlm-chat-7b`, etc.                                                                                                                    | ✅︎                     | ✅︎                          | ✅︎                     |
 | `InternLM2ForCausalLM`                            | InternLM2                                           | `internlm/internlm2-7b`, `internlm/internlm2-chat-7b`, etc.                                                                                                                  | ✅︎                     | ✅︎                          | ✅︎                     |
 | `InternLM3ForCausalLM`                            | InternLM3                                           | `internlm/internlm3-8b-instruct`, etc.                                                                                                                                       | ✅︎                     | ✅︎                          | ✅︎                     |
@@ -386,7 +390,7 @@ Specified using `--task generate`.
 | `TeleChat2ForCausalLM`                            | TeleChat2                                           | `Tele-AI/TeleChat2-3B`, `Tele-AI/TeleChat2-7B`, `Tele-AI/TeleChat2-35B`, etc.                                                                                                | ✅︎                     | ✅︎                          | ✅︎                     |
 | `TeleFLMForCausalLM`                              | TeleFLM                                             | `CofeAI/FLM-2-52B-Instruct-2407`, `CofeAI/Tele-FLM`, etc.                                                                                                                    | ✅︎                     | ✅︎                          | ✅︎                     |
 | `XverseForCausalLM`                               | XVERSE                                              | `xverse/XVERSE-7B-Chat`, `xverse/XVERSE-13B-Chat`, `xverse/XVERSE-65B-Chat`, etc.                                                                                            | ✅︎                     | ✅︎                          | ✅︎                     |
-| `MiniMaxM1ForCausalLM`                        | MiniMax-Text                                        | `MiniMaxAI/MiniMax-M1-40k`, `MiniMaxAI/MiniMax-M1-80k`etc.                                                                                                                                  |                        |                             |                       |
+| `MiniMaxM1ForCausalLM`                        | MiniMax-Text                                        | `MiniMaxAI/MiniMax-M1-40k`, `MiniMaxAI/MiniMax-M1-80k`etc.                                                                                                                       |                        |                             |                       |
 | `MiniMaxText01ForCausalLM`                        | MiniMax-Text                                        | `MiniMaxAI/MiniMax-Text-01`, etc.                                                                                                                                            |                        |                             |                       |
 | `Zamba2ForCausalLM`                               | Zamba2                                              | `Zyphra/Zamba2-7B-instruct`, `Zyphra/Zamba2-2.7B-instruct`, `Zyphra/Zamba2-1.2B-instruct`, etc.                                                                              |                        |                             |                       |
 
@@ -552,10 +556,12 @@ Specified using `--task generate`.
 | `FuyuForCausalLM`                            | Fuyu                                                                     | T + I                                                                 | `adept/fuyu-8b` etc.                                                                                                                                    |                        | ✅︎                          | ✅︎                    |
 | `Gemma3ForConditionalGeneration`             | Gemma 3                                                                  | T + I<sup>+</sup>                                                     | `google/gemma-3-4b-it`, `google/gemma-3-27b-it`, etc.                                                                                                   | ✅︎                     | ✅︎                          | ⚠️                    |
 | `GLM4VForCausalLM`<sup>^</sup>               | GLM-4V                                                                   | T + I                                                                 | `THUDM/glm-4v-9b`, `THUDM/cogagent-9b-20241220` etc.                                                                                                    | ✅︎                     | ✅︎                          | ✅︎                    |
+| `Glm4vForConditionalGeneration`              | GLM-4.1V-Thinking                                                        | T + I<sup>E+</sup> + V<sup>E+</sup>                                   | `THUDM/GLM-4.1V-9B-Thinkg`,  etc.                                                                                                                       | ✅︎                     | ✅︎                          | ✅︎                    |
 | `GraniteSpeechForConditionalGeneration`      | Granite Speech                                                           | T + A                                                                 | `ibm-granite/granite-speech-3.3-8b`                                                                                                                     | ✅︎                     | ✅︎                          | ✅︎                    |
 | `H2OVLChatModel`                             | H2OVL                                                                    | T + I<sup>E+</sup>                                                    | `h2oai/h2ovl-mississippi-800m`, `h2oai/h2ovl-mississippi-2b`, etc.                                                                                      |                        | ✅︎                          | ✅︎\*                  |
 | `Idefics3ForConditionalGeneration`           | Idefics3                                                                 | T + I                                                                 | `HuggingFaceM4/Idefics3-8B-Llama3` etc.                                                                                                                 | ✅︎                     |                             |  ✅︎                   |
 | `InternVLChatModel`                          | InternVL 3.0, InternVideo 2.5, InternVL 2.5, Mono-InternVL, InternVL 2.0 | T + I<sup>E+</sup> + (V<sup>E+</sup>)                                 | `OpenGVLab/InternVL3-9B`, `OpenGVLab/InternVideo2_5_Chat_8B`, `OpenGVLab/InternVL2_5-4B`, `OpenGVLab/Mono-InternVL-2B`, `OpenGVLab/InternVL2-4B`, etc.  | ✅︎                     | ✅︎                          | ✅︎                    |
+| `KeyeForConditionalGeneration`               | Keye-VL-8B-Preview                                                       | T + I<sup>E+</sup> + V<sup>E+</sup>                                   | `Kwai-Keye/Keye-VL-8B-Preview`                                                                                                                          |                        |                             | ✅︎                    |
 | `KimiVLForConditionalGeneration`             | Kimi-VL-A3B-Instruct, Kimi-VL-A3B-Thinking                               | T + I<sup>+</sup>                                                     | `moonshotai/Kimi-VL-A3B-Instruct`, `moonshotai/Kimi-VL-A3B-Thinking`                                                                                    |                        |                             | ✅︎                    |
 | `Llama4ForConditionalGeneration`             | Llama 4                                                                  | T + I<sup>+</sup>                                                     | `meta-llama/Llama-4-Scout-17B-16E-Instruct`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct`, etc. |                        | ✅︎                          | ✅︎                    |
 | `LlavaForConditionalGeneration`              | LLaVA-1.5                                                                | T + I<sup>E+</sup>                                                    | `llava-hf/llava-1.5-7b-hf`, `TIGER-Lab/Mantis-8B-siglip-llama3` (see note), etc.                                                                        |                        | ✅︎                          | ✅︎                    |
@@ -582,7 +588,7 @@ Specified using `--task generate`.
 | `SkyworkR1VChatModel`                        | Skywork-R1V-38B                                                          | T + I                                                                 | `Skywork/Skywork-R1V-38B`                                                                                                                               |                        | ✅︎                          | ✅︎                    |
 | `SmolVLMForConditionalGeneration`            | SmolVLM2                                                                 | T + I                                                                 | `SmolVLM2-2.2B-Instruct`                                                                                                                                | ✅︎                     |                             | ✅︎                    |
 | `TarsierForConditionalGeneration`            | Tarsier                                                                  | T + I<sup>E+</sup>                                                    | `omni-search/Tarsier-7b`,`omni-search/Tarsier-34b`                                                                                                      |                        | ✅︎                          | ✅︎                    |
-| `Tarsier2ForConditionalGeneration`<sup>^</sup>            | Tarsier2                                                                  | T + I<sup>E+</sup> + V<sup>E+</sup>                                                    | `omni-research/Tarsier2-Recap-7b`,`omni-research/Tarsier2-7b-0115`                                                                                                      |                        | ✅︎                          | ✅︎                    |
+| `Tarsier2ForConditionalGeneration`<sup>^</sup> | Tarsier2                                                                 | T + I<sup>E+</sup> + V<sup>E+</sup>                                                | `omni-research/Tarsier2-Recap-7b`,`omni-research/Tarsier2-7b-0115`                                                                                      |                        | ✅︎                          | ✅︎                    |
 
 <sup>^</sup> You need to set the architecture name via `--hf-overrides` to match the one in vLLM.  
 &nbsp;&nbsp;&nbsp;&nbsp;• For example, to use DeepSeek-VL2 series models:  

docs/serving/distributed_serving.md

@@ -100,7 +100,50 @@ vllm serve /path/to/the/model/in/the/container \
      --tensor-parallel-size 16
 ```
 
-To make tensor parallel performant, you should make sure the communication between nodes is efficient, e.g. using high-speed network cards like Infiniband. To correctly set up the cluster to use Infiniband, append additional arguments like `--privileged -e NCCL_IB_HCA=mlx5` to the `run_cluster.sh` script. Please contact your system administrator for more information on how to set up the flags. One way to confirm if the Infiniband is working is to run vLLM with `NCCL_DEBUG=TRACE` environment variable set, e.g. `NCCL_DEBUG=TRACE vllm serve ...` and check the logs for the NCCL version and the network used. If you find `[send] via NET/Socket` in the logs, it means NCCL uses raw TCP Socket, which is not efficient for cross-node tensor parallel. If you find `[send] via NET/IB/GDRDMA` in the logs, it means NCCL uses Infiniband with GPU-Direct RDMA, which is efficient.
+To make tensor parallel performant, you should make sure the communication between nodes is efficient, e.g. using high-speed network cards like InfiniBand. To correctly set up the cluster to use InfiniBand, append additional arguments like `--privileged -e NCCL_IB_HCA=mlx5` to the `run_cluster.sh` script. Please contact your system administrator for more information on how to set up the flags. One way to confirm if the InfiniBand is working is to run vLLM with `NCCL_DEBUG=TRACE` environment variable set, e.g. `NCCL_DEBUG=TRACE vllm serve ...` and check the logs for the NCCL version and the network used. If you find `[send] via NET/Socket` in the logs, it means NCCL uses raw TCP Socket, which is not efficient for cross-node tensor parallel. If you find `[send] via NET/IB/GDRDMA` in the logs, it means NCCL uses InfiniBand with GPUDirect RDMA, which is efficient.
+
+### GPUDirect RDMA
+
+To enable GPUDirect RDMA with vLLM, specific configuration tweaks are needed. This setup ensures:
+
+- `IPC_LOCK` Security Context: Add the `IPC_LOCK` capability to the container’s security context to lock memory pages and prevent swapping to disk.
+- Shared Memory with `/dev/shm`: Mount `/dev/shm` in the pod spec to provide shared memory for IPC.
+
+When using Docker, you can set up the container as follows:
+
+```bash
+docker run --gpus all \
+    --ipc=host \
+    --shm-size=16G \
+    -v /dev/shm:/dev/shm \
+    vllm/vllm-openai
+```
+
+When using Kubernetes, you can set up the pod spec as follows:
+
+```yaml
+...
+spec:
+  containers:
+    - name: vllm
+      image: vllm/vllm-openai
+      securityContext:
+        capabilities:
+          add: ["IPC_LOCK"]
+      volumeMounts:
+        - mountPath: /dev/shm
+          name: dshm
+      resources:
+        limits:
+          nvidia.com/gpu: 8
+        requests:
+          nvidia.com/gpu: 8
+  volumes:
+    - name: dshm
+      emptyDir:
+        medium: Memory
+...
+```
 
 !!! warning
     After you start the Ray cluster, you'd better also check the GPU-GPU communication between nodes. It can be non-trivial to set up. Please refer to the [sanity check script][troubleshooting-incorrect-hardware-driver] for more information. If you need to set some environment variables for the communication configuration, you can append them to the `run_cluster.sh` script, e.g. `-e NCCL_SOCKET_IFNAME=eth0`. Note that setting environment variables in the shell (e.g. `NCCL_SOCKET_IFNAME=eth0 vllm serve ...`) only works for the processes in the same node, not for the processes in the other nodes. Setting environment variables when you create the cluster is the recommended way. See <gh-issue:6803> for more information.

docs/serving/openai_compatible_server.md

@@ -1,7 +1,7 @@
 ---
 title: OpenAI-Compatible Server
 ---
-[](){ #openai-compatible-server }
+[](){ #serving-openai-compatible-server }
 
 vLLM provides an HTTP server that implements OpenAI's [Completions API](https://platform.openai.com/docs/api-reference/completions), [Chat API](https://platform.openai.com/docs/api-reference/chat), and more! This functionality lets you serve models and interact with them using an HTTP client.
 

docs/usage/troubleshooting.md

@@ -273,6 +273,27 @@ But you are sure that the model is in the [list of supported models][supported-m
 
 If you see an error like `RuntimeError: Failed to infer device type`, it means that vLLM failed to infer the device type of the runtime environment. You can check [the code](gh-file:vllm/platforms/__init__.py) to see how vLLM infers the device type and why it is not working as expected. After [this PR](gh-pr:14195), you can also set the environment variable `VLLM_LOGGING_LEVEL=DEBUG` to see more detailed logs to help debug the issue.
 
+## NCCL error: unhandled system error during `ncclCommInitRank`
+
+If your serving workload uses GPUDirect RDMA for distributed serving across multiple nodes and encounters an error during `ncclCommInitRank`, with no clear error message even with `NCCL_DEBUG=INFO` set, it might look like this:
+
+```text
+Error executing method 'init_device'. This might cause deadlock in distributed execution.
+Traceback (most recent call last):
+...
+   File "/usr/local/lib/python3.12/dist-packages/vllm/distributed/device_communicators/pynccl.py", line 99, in __init__
+     self.comm: ncclComm_t = self.nccl.ncclCommInitRank(
+                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^
+   File "/usr/local/lib/python3.12/dist-packages/vllm/distributed/device_communicators/pynccl_wrapper.py", line 277, in ncclCommInitRank
+     self.NCCL_CHECK(self._funcs["ncclCommInitRank"](ctypes.byref(comm),
+   File "/usr/local/lib/python3.12/dist-packages/vllm/distributed/device_communicators/pynccl_wrapper.py", line 256, in NCCL_CHECK
+     raise RuntimeError(f"NCCL error: {error_str}")
+ RuntimeError: NCCL error: unhandled system error (run with NCCL_DEBUG=INFO for details)
+...
+```
+
+This indicates vLLM failed to initialize the NCCL communicator, possibly due to a missing `IPC_LOCK` linux capability  or an unmounted `/dev/shm`. Refer to [Distributed Inference and Serving](../serving/distributed_serving.md#running-vllm-on-multiple-nodes) for guidance on properly configuring the environment for distributed serving.
+
 ## Known Issues
 
 - In `v0.5.2`, `v0.5.3`, and `v0.5.3.post1`, there is a bug caused by [zmq](https://github.com/zeromq/pyzmq/issues/2000) , which can occasionally cause vLLM to hang depending on the machine configuration. The solution is to upgrade to the latest version of `vllm` to include the [fix](gh-pr:6759).

examples/offline_inference/data_parallel.py

@@ -64,6 +64,18 @@ def parse_args():
     parser.add_argument(
         "--trust-remote-code", action="store_true", help="Trust remote code."
     )
+    parser.add_argument(
+        "--max-num-seqs",
+        type=int,
+        default=64,
+        help=("Maximum number of sequences to be processed in a single iteration."),
+    )
+    parser.add_argument(
+        "--gpu-memory-utilization",
+        type=float,
+        default=0.8,
+        help=("Fraction of GPU memory vLLM is allowed to allocate (0.0, 1.0]."),
+    )
     return parser.parse_args()
 
 
@@ -77,6 +89,8 @@ def main(
     GPUs_per_dp_rank,
     enforce_eager,
     trust_remote_code,
+    max_num_seqs,
+    gpu_memory_utilization,
 ):
     os.environ["VLLM_DP_RANK"] = str(global_dp_rank)
     os.environ["VLLM_DP_RANK_LOCAL"] = str(local_dp_rank)
@@ -127,6 +141,8 @@ def main(
         enforce_eager=enforce_eager,
         enable_expert_parallel=True,
         trust_remote_code=trust_remote_code,
+        max_num_seqs=max_num_seqs,
+        gpu_memory_utilization=gpu_memory_utilization,
     )
     outputs = llm.generate(prompts, sampling_params)
     # Print the outputs.
@@ -181,6 +197,8 @@ if __name__ == "__main__":
                 tp_size,
                 args.enforce_eager,
                 args.trust_remote_code,
+                args.max_num_seqs,
+                args.gpu_memory_utilization,
             ),
         )
         proc.start()

examples/offline_inference/eagle.py

@@ -1,144 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import argparse
-import json
-import os
-
-from transformers import AutoTokenizer
-
-from vllm import LLM, SamplingParams
-from vllm.v1.metrics.reader import Counter, Vector
-
-
-def load_prompts(dataset_path, num_prompts):
-    if os.path.exists(dataset_path):
-        prompts = []
-        try:
-            with open(dataset_path) as f:
-                for line in f:
-                    data = json.loads(line)
-                    prompts.append(data["turns"][0])
-        except Exception as e:
-            print(f"Error reading dataset: {e}")
-            return []
-    else:
-        prompts = ["The future of AI is", "The president of the United States is"]
-
-    return prompts[:num_prompts]
-
-
-def parse_args():
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--dataset",
-        type=str,
-        default="./examples/data/gsm8k.jsonl",
-        help="downloaded from the eagle repo "
-        "https://github.com/SafeAILab/EAGLE/blob/main/eagle/data/",
-    )
-    parser.add_argument(
-        "--method", type=str, default="eagle", choices=["eagle", "eagle3"]
-    )
-    parser.add_argument("--max_num_seqs", type=int, default=8)
-    parser.add_argument("--num_prompts", type=int, default=80)
-    parser.add_argument("--num_spec_tokens", type=int, default=2)
-    parser.add_argument("--tp", type=int, default=1)
-    parser.add_argument("--draft_tp", type=int, default=1)
-    parser.add_argument("--enforce_eager", action="store_true")
-    parser.add_argument("--enable_chunked_prefill", action="store_true")
-    parser.add_argument("--max_num_batched_tokens", type=int, default=2048)
-    parser.add_argument("--temp", type=float, default=0)
-    return parser.parse_args()
-
-
-def main():
-    args = parse_args()
-
-    model_dir = "meta-llama/Llama-3.1-8B-Instruct"
-
-    if args.method == "eagle":
-        eagle_dir = "yuhuili/EAGLE-LLaMA3.1-Instruct-8B"
-    elif args.method == "eagle3":
-        eagle_dir = "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B"
-    else:
-        raise ValueError(f"unknown method: {args.method}")
-
-    max_model_len = 2048
-
-    tokenizer = AutoTokenizer.from_pretrained(model_dir)
-
-    prompts = load_prompts(args.dataset, args.num_prompts)
-
-    prompt_ids = [
-        tokenizer.apply_chat_template(
-            [{"role": "user", "content": prompt}], add_generation_prompt=True
-        )
-        for prompt in prompts
-    ]
-
-    llm = LLM(
-        model=model_dir,
-        trust_remote_code=True,
-        tensor_parallel_size=args.tp,
-        enable_chunked_prefill=args.enable_chunked_prefill,
-        max_num_batched_tokens=args.max_num_batched_tokens,
-        enforce_eager=args.enforce_eager,
-        max_model_len=max_model_len,
-        max_num_seqs=args.max_num_seqs,
-        gpu_memory_utilization=0.8,
-        speculative_config={
-            "method": args.method,
-            "model": eagle_dir,
-            "num_speculative_tokens": args.num_spec_tokens,
-            "draft_tensor_parallel_size": args.draft_tp,
-            "max_model_len": max_model_len,
-        },
-        disable_log_stats=False,
-    )
-
-    sampling_params = SamplingParams(temperature=args.temp, max_tokens=256)
-
-    outputs = llm.generate(prompt_token_ids=prompt_ids, sampling_params=sampling_params)
-
-    # print the generated text
-    for output in outputs:
-        print("-" * 50)
-        print(f"prompt: {output.prompt}")
-        print(f"generated text: {output.outputs[0].text}")
-        print("-" * 50)
-
-    try:
-        metrics = llm.get_metrics()
-    except AssertionError:
-        print("Metrics are not supported in the V0 engine.")
-        return
-
-    num_drafts = num_accepted = 0
-    acceptance_counts = [0] * args.num_spec_tokens
-    for metric in metrics:
-        if metric.name == "vllm:spec_decode_num_drafts":
-            assert isinstance(metric, Counter)
-            num_drafts += metric.value
-        elif metric.name == "vllm:spec_decode_num_accepted_tokens":
-            assert isinstance(metric, Counter)
-            num_accepted += metric.value
-        elif metric.name == "vllm:spec_decode_num_accepted_tokens_per_pos":
-            assert isinstance(metric, Vector)
-            for pos in range(len(metric.values)):
-                acceptance_counts[pos] += metric.values[pos]
-
-    print("-" * 50)
-    print(f"mean acceptance length: {1 + (num_accepted / num_drafts):.2f}")
-    print("-" * 50)
-
-    # print acceptance at each token position
-    for i in range(len(acceptance_counts)):
-        print(f"acceptance at token {i}:{acceptance_counts[i] / num_drafts:.2f}")
-
-
-if __name__ == "__main__":
-    print(
-        "[WARNING] Use examples/offline_inference/spec_decode.py"
-        " instead of this script."
-    )
-    main()

examples/offline_inference/spec_decode.py

@@ -16,24 +16,17 @@ def parse_args():
     parser = FlexibleArgumentParser()
     add_dataset_parser(parser)
     parser.add_argument(
-        "--dataset",
+        "--method",
         type=str,
-        default="./examples/data/gsm8k.jsonl",
-        help="downloaded from the eagle repo "
-        "https://github.com/SafeAILab/EAGLE/blob/main/eagle/data/",
+        default="eagle",
+        choices=["ngram", "eagle", "eagle3", "mtp"],
     )
-    parser.add_argument(
-        "--method", type=str, default="eagle", choices=["ngram", "eagle", "eagle3"]
-    )
-    parser.add_argument("--max-num-seqs", type=int, default=8)
     parser.add_argument("--num-spec-tokens", type=int, default=2)
     parser.add_argument("--prompt-lookup-max", type=int, default=5)
     parser.add_argument("--prompt-lookup-min", type=int, default=2)
     parser.add_argument("--tp", type=int, default=1)
-    parser.add_argument("--draft-tp", type=int, default=1)
     parser.add_argument("--enforce-eager", action="store_true")
     parser.add_argument("--enable-chunked-prefill", action="store_true")
-    parser.add_argument("--max-num-batched-tokens", type=int, default=2048)
     parser.add_argument("--temp", type=float, default=0)
     parser.add_argument("--top-p", type=float, default=1.0)
     parser.add_argument("--top-k", type=int, default=-1)
@@ -41,7 +34,6 @@ def parse_args():
     parser.add_argument("--output-len", type=int, default=256)
     parser.add_argument("--model-dir", type=str, default=None)
     parser.add_argument("--eagle-dir", type=str, default=None)
-    parser.add_argument("--max-model-len", type=int, default=2048)
     return parser.parse_args()
 
 
@@ -71,8 +63,6 @@ def main():
             "method": args.method,
             "model": eagle_dir,
             "num_speculative_tokens": args.num_spec_tokens,
-            "draft_tensor_parallel_size": args.draft_tp,
-            "max_model_len": args.max_model_len,
         }
     elif args.method == "ngram":
         speculative_config = {
@@ -80,7 +70,6 @@ def main():
             "num_speculative_tokens": args.num_spec_tokens,
             "prompt_lookup_max": args.prompt_lookup_max,
             "prompt_lookup_min": args.prompt_lookup_min,
-            "max_model_len": args.max_model_len,
         }
     else:
         raise ValueError(f"unknown method: {args.method}")
@@ -90,10 +79,7 @@ def main():
         trust_remote_code=True,
         tensor_parallel_size=args.tp,
         enable_chunked_prefill=args.enable_chunked_prefill,
-        max_num_batched_tokens=args.max_num_batched_tokens,
         enforce_eager=args.enforce_eager,
-        max_model_len=args.max_model_len,
-        max_num_seqs=args.max_num_seqs,
         gpu_memory_utilization=0.8,
         speculative_config=speculative_config,
         disable_log_stats=False,
@@ -116,27 +102,41 @@ def main():
         print("Metrics are not supported in the V0 engine.")
         return
 
-    num_drafts = num_accepted = 0
+    total_num_output_tokens = sum(
+        len(output.outputs[0].token_ids) for output in outputs
+    )
+    num_drafts = 0
+    num_draft_tokens = 0
+    num_accepted_tokens = 0
     acceptance_counts = [0] * args.num_spec_tokens
     for metric in metrics:
         if metric.name == "vllm:spec_decode_num_drafts":
             assert isinstance(metric, Counter)
             num_drafts += metric.value
+        elif metric.name == "vllm:spec_decode_num_draft_tokens":
+            assert isinstance(metric, Counter)
+            num_draft_tokens += metric.value
         elif metric.name == "vllm:spec_decode_num_accepted_tokens":
             assert isinstance(metric, Counter)
-            num_accepted += metric.value
+            num_accepted_tokens += metric.value
         elif metric.name == "vllm:spec_decode_num_accepted_tokens_per_pos":
             assert isinstance(metric, Vector)
             for pos in range(len(metric.values)):
                 acceptance_counts[pos] += metric.values[pos]
 
     print("-" * 50)
-    print(f"mean acceptance length: {1 + (num_accepted / num_drafts):.2f}")
+    print(f"total_num_output_tokens: {total_num_output_tokens}")
+    print(f"num_drafts: {num_drafts}")
+    print(f"num_draft_tokens: {num_draft_tokens}")
+    print(f"num_accepted_tokens: {num_accepted_tokens}")
+    acceptance_length = 1 + (num_accepted_tokens / num_drafts) if num_drafts > 0 else 1
+    print(f"mean acceptance length: {acceptance_length:.2f}")
     print("-" * 50)
 
     # print acceptance at each token position
     for i in range(len(acceptance_counts)):
-        print(f"acceptance at token {i}:{acceptance_counts[i] / num_drafts:.2f}")
+        acceptance_rate = acceptance_counts[i] / num_drafts if num_drafts > 0 else 0
+        print(f"acceptance at token {i}: {acceptance_rate:.2f}")
 
 
 if __name__ == "__main__":

examples/offline_inference/vision_language.py

@@ -248,6 +248,42 @@ def run_glm4v(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
+# GLM-4.1V
+def run_glm4_1v(questions: list[str], modality: str) -> ModelRequestData:
+    model_name = "THUDM/GLM-4.1V-9B-Thinking"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=4096,
+        max_num_seqs=2,
+        mm_processor_kwargs={
+            "size": {"shortest_edge": 12544, "longest_edge": 47040000},
+            "fps": 1,
+        },
+        limit_mm_per_prompt={modality: 1},
+        enforce_eager=True,
+    )
+
+    if modality == "image":
+        placeholder = "<|begin_of_image|><|image|><|end_of_image|>"
+    elif modality == "video":
+        placeholder = "<|begin_of_video|><|video|><|end_of_video|>"
+
+    prompts = [
+        (
+            "[gMASK]<sop><|system|>\nYou are a helpful assistant.<|user|>\n"
+            f"{placeholder}"
+            f"{question}<|assistant|>assistant\n"
+        )
+        for question in questions
+    ]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
+
 # H2OVL-Mississippi
 def run_h2ovl(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -393,6 +429,37 @@ def run_internvl(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
+# Keye-VL
+def run_keye_vl(questions: list[str], modality: str) -> ModelRequestData:
+    model_name = "Kwai-Keye/Keye-VL-8B-Preview"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=8192,
+        trust_remote_code=True,
+        limit_mm_per_prompt={modality: 1},
+    )
+
+    if modality == "image":
+        placeholder = "<|image_pad|>"
+    elif modality == "video":
+        placeholder = "<|video_pad|>"
+
+    prompts = [
+        (
+            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
+            f"{question}<|im_end|>\n"
+            "<|im_start|>assistant\n"
+        )
+        for question in questions
+    ]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
+
 # Kimi-VL
 def run_kimi_vl(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -1114,9 +1181,11 @@ model_example_map = {
     "fuyu": run_fuyu,
     "gemma3": run_gemma3,
     "glm4v": run_glm4v,
+    "glm4_1v": run_glm4_1v,
     "h2ovl_chat": run_h2ovl,
     "idefics3": run_idefics3,
     "internvl_chat": run_internvl,
+    "keye_vl": run_keye_vl,
     "kimi_vl": run_kimi_vl,
     "llava": run_llava,
     "llava-next": run_llava_next,
@@ -1172,10 +1241,11 @@ def get_multi_modal_input(args):
     if args.modality == "video":
         # Input video and question
         video = VideoAsset(name="baby_reading", num_frames=args.num_frames).np_ndarrays
+        metadata = VideoAsset(name="baby_reading", num_frames=args.num_frames).metadata
         vid_questions = ["Why is this video funny?"]
 
         return {
-            "data": video,
+            "data": [(video, metadata)] if args.model_type == "glm4_1v" else video,
             "questions": vid_questions,
         }
 

examples/offline_inference/vision_language_multi_image.py

@@ -423,6 +423,43 @@ def load_llama4(question: str, image_urls: list[str]) -> ModelRequestData:
     )
 
 
+def load_keye_vl(question: str, image_urls: list[str]) -> ModelRequestData:
+    model_name = "Kwai-Keye/Keye-VL-8B-Preview"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        trust_remote_code=True,
+        max_model_len=8192,
+        max_num_seqs=5,
+        limit_mm_per_prompt={"image": len(image_urls)},
+    )
+
+    placeholders = [{"type": "image", "image": url} for url in image_urls]
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                *placeholders,
+                {"type": "text", "text": question},
+            ],
+        },
+    ]
+
+    processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
+
+    prompt = processor.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+
+    image_data = [fetch_image(url) for url in image_urls]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+        image_data=image_data,
+    )
+
+
 def load_kimi_vl(question: str, image_urls: list[str]) -> ModelRequestData:
     model_name = "moonshotai/Kimi-VL-A3B-Instruct"
 
@@ -862,6 +899,7 @@ model_example_map = {
     "h2ovl_chat": load_h2ovl,
     "idefics3": load_idefics3,
     "internvl_chat": load_internvl,
+    "keye_vl": load_keye_vl,
     "kimi_vl": load_kimi_vl,
     "llava": load_llava,
     "llava-next": load_llava_next,

examples/online_serving/disaggregated_serving_p2p_nccl_xpyd/disagg_example_p2p_nccl_xpyd.sh

@@ -0,0 +1,245 @@
+#!/bin/bash
+
+# =============================================================================
+# vLLM Disaggregated Serving Script - P2P NCCL XpYd Architecture
+# =============================================================================
+# This script demonstrates disaggregated prefill and decode serving using
+# P2P NCCL communication. The architecture supports various XpYd configurations:
+#
+# - 1P3D: 1 Prefill server + 3 Decode servers (current default)
+# - 3P1D: 3 Prefill servers + 1 Decode server
+# - etc.
+#
+# Configuration can be customized via environment variables:
+#   MODEL: Model to serve
+#   PREFILL_GPUS: Comma-separated GPU IDs for prefill servers
+#   DECODE_GPUS: Comma-separated GPU IDs for decode servers
+#   PREFILL_PORTS: Comma-separated ports for prefill servers
+#   DECODE_PORTS: Comma-separated ports for decode servers
+#   PROXY_PORT: Proxy server port used to setup XpYd connection.
+#   TIMEOUT_SECONDS: Server startup timeout
+# =============================================================================
+
+# Configuration - can be overridden via environment variables
+MODEL=${MODEL:-meta-llama/Llama-3.1-8B-Instruct}
+TIMEOUT_SECONDS=${TIMEOUT_SECONDS:-1200}
+PROXY_PORT=${PROXY_PORT:-30001}
+
+# Default 1P3D configuration (1 Prefill + 3 Decode)
+PREFILL_GPUS=${PREFILL_GPUS:-0}
+DECODE_GPUS=${DECODE_GPUS:-1,2,3}
+PREFILL_PORTS=${PREFILL_PORTS:-20003}
+DECODE_PORTS=${DECODE_PORTS:-20005,20007,20009} 
+
+echo "Warning: P2P NCCL disaggregated prefill XpYd support for vLLM v1 is experimental and subject to change."
+echo ""
+echo "Architecture Configuration:"
+echo "  Model: $MODEL"
+echo "  Prefill GPUs: $PREFILL_GPUS, Ports: $PREFILL_PORTS"
+echo "  Decode GPUs: $DECODE_GPUS, Ports: $DECODE_PORTS"
+echo "  Proxy Port: $PROXY_PORT"
+echo "  Timeout: ${TIMEOUT_SECONDS}s"
+echo ""
+
+PIDS=()
+
+# Switch to the directory of the current script
+cd "$(dirname "${BASH_SOURCE[0]}")"
+
+check_required_files() {
+    local files=("disagg_proxy_p2p_nccl_xpyd.py")
+    for file in "${files[@]}"; do
+        if [[ ! -f "$file" ]]; then
+            echo "Required file $file not found in $(pwd)"
+            exit 1
+        fi
+    done
+}
+
+check_hf_token() {
+    if [ -z "$HF_TOKEN" ]; then
+        echo "HF_TOKEN is not set. Please set it to your Hugging Face token."
+        echo "Example: export HF_TOKEN=your_token_here"
+        exit 1
+    fi
+    if [[ "$HF_TOKEN" != hf_* ]]; then
+        echo "HF_TOKEN is not a valid Hugging Face token. Please set it to your Hugging Face token."
+        exit 1
+    fi
+    echo "HF_TOKEN is set and valid."
+}
+
+check_num_gpus() {
+    # Check if the number of GPUs are >=2 via nvidia-smi
+    num_gpus=$(nvidia-smi --query-gpu=name --format=csv,noheader | wc -l)
+    if [ "$num_gpus" -lt 2 ]; then
+        echo "You need at least 2 GPUs to run disaggregated prefill."
+        exit 1
+    else
+        echo "Found $num_gpus GPUs."
+    fi
+}
+
+ensure_python_library_installed() {
+    echo "Checking if $1 is installed..."
+    if ! python3 -c "import $1" > /dev/null 2>&1; then
+        echo "$1 is not installed. Please install it via pip install $1."
+        exit 1
+    else
+        echo "$1 is installed."
+    fi
+}
+
+cleanup() {
+    echo "Stopping everything…"
+    trap - INT TERM        # prevent re-entrancy
+    kill -- -$$            # negative PID  ==  "this whole process-group"
+    wait                   # reap children so we don't leave zombies
+    exit 0
+}
+
+wait_for_server() {
+  local port=$1
+  local timeout_seconds=$TIMEOUT_SECONDS
+  local start_time=$(date +%s)
+
+  echo "Waiting for server on port $port..."
+
+  while true; do
+    if curl -s "localhost:${port}/v1/completions" > /dev/null; then
+      echo "Server on port $port is ready."
+      return 0
+    fi
+
+    local now=$(date +%s)
+    if (( now - start_time >= timeout_seconds )); then
+      echo "Timeout waiting for server on port $port"
+      return 1
+    fi
+
+    sleep 1
+  done
+}
+
+main() {
+    check_required_files
+    check_hf_token
+    check_num_gpus
+    ensure_python_library_installed pandas
+    ensure_python_library_installed datasets
+    ensure_python_library_installed vllm
+    ensure_python_library_installed quart
+
+    trap cleanup INT
+    trap cleanup USR1
+    trap cleanup TERM
+
+    echo "Launching disaggregated serving components..."
+    echo "Please check the log files for detailed output:"
+    echo "  - prefill*.log: Prefill server logs"
+    echo "  - decode*.log: Decode server logs"
+    echo "  - proxy.log: Proxy server log"
+
+    # =============================================================================
+    # Launch Proxy Server
+    # =============================================================================
+    echo ""
+    echo "Starting proxy server on port $PROXY_PORT..."
+    python3 disagg_proxy_p2p_nccl_xpyd.py &
+    PIDS+=($!)
+
+    # Parse GPU and port arrays
+    IFS=',' read -ra PREFILL_GPU_ARRAY <<< "$PREFILL_GPUS"
+    IFS=',' read -ra DECODE_GPU_ARRAY <<< "$DECODE_GPUS"
+    IFS=',' read -ra PREFILL_PORT_ARRAY <<< "$PREFILL_PORTS"
+    IFS=',' read -ra DECODE_PORT_ARRAY <<< "$DECODE_PORTS"
+
+    # =============================================================================
+    # Launch Prefill Servers (X Producers)
+    # =============================================================================
+    echo ""
+    echo "Starting ${#PREFILL_GPU_ARRAY[@]} prefill server(s)..."
+    for i in "${!PREFILL_GPU_ARRAY[@]}"; do
+        local gpu_id=${PREFILL_GPU_ARRAY[$i]}
+        local port=${PREFILL_PORT_ARRAY[$i]}
+        local kv_port=$((21001 + i))
+        
+        echo "  Prefill server $((i+1)): GPU $gpu_id, Port $port, KV Port $kv_port"
+        CUDA_VISIBLE_DEVICES=$gpu_id VLLM_USE_V1=1 vllm serve $MODEL \
+        --enforce-eager \
+        --host 0.0.0.0 \
+        --port $port \
+        --tensor-parallel-size 1 \
+        --seed 1024 \
+        --dtype float16 \
+        --max-model-len 10000 \
+        --max-num-batched-tokens 10000 \
+        --max-num-seqs 256 \
+        --trust-remote-code \
+        --gpu-memory-utilization 0.9 \
+        --disable-log-request \
+        --kv-transfer-config \
+        "{\"kv_connector\":\"P2pNcclConnector\",\"kv_role\":\"kv_producer\",\"kv_buffer_size\":\"1e1\",\"kv_port\":\"$kv_port\",\"kv_connector_extra_config\":{\"proxy_ip\":\"0.0.0.0\",\"proxy_port\":\"$PROXY_PORT\",\"http_port\":\"$port\",\"send_type\":\"PUT_ASYNC\",\"nccl_num_channels\":\"16\"}}" > prefill$((i+1)).log 2>&1 &
+        PIDS+=($!)
+    done
+
+    # =============================================================================
+    # Launch Decode Servers (Y Decoders)
+    # =============================================================================
+    echo ""
+    echo "Starting ${#DECODE_GPU_ARRAY[@]} decode server(s)..."
+    for i in "${!DECODE_GPU_ARRAY[@]}"; do
+        local gpu_id=${DECODE_GPU_ARRAY[$i]}
+        local port=${DECODE_PORT_ARRAY[$i]}
+        local kv_port=$((22001 + i))
+        
+        echo "  Decode server $((i+1)): GPU $gpu_id, Port $port, KV Port $kv_port"
+        VLLM_USE_V1=1 CUDA_VISIBLE_DEVICES=$gpu_id vllm serve $MODEL \
+        --enforce-eager \
+        --host 0.0.0.0 \
+        --port $port \
+        --tensor-parallel-size 1 \
+        --seed 1024 \
+        --dtype float16 \
+        --max-model-len 10000 \
+        --max-num-batched-tokens 10000 \
+        --max-num-seqs 256 \
+        --trust-remote-code \
+        --gpu-memory-utilization 0.7 \
+        --disable-log-request \
+        --kv-transfer-config \
+        "{\"kv_connector\":\"P2pNcclConnector\",\"kv_role\":\"kv_consumer\",\"kv_buffer_size\":\"8e9\",\"kv_port\":\"$kv_port\",\"kv_connector_extra_config\":{\"proxy_ip\":\"0.0.0.0\",\"proxy_port\":\"$PROXY_PORT\",\"http_port\":\"$port\",\"send_type\":\"PUT_ASYNC\",\"nccl_num_channels\":\"16\"}}" > decode$((i+1)).log 2>&1 &
+        PIDS+=($!)
+    done
+
+    # =============================================================================
+    # Wait for All Servers to Start
+    # =============================================================================
+    echo ""
+    echo "Waiting for all servers to start..."
+    for port in "${PREFILL_PORT_ARRAY[@]}" "${DECODE_PORT_ARRAY[@]}"; do
+        if ! wait_for_server $port; then
+            echo "Failed to start server on port $port"
+            cleanup
+            exit 1
+        fi
+    done
+
+    echo ""
+    echo "All servers are up. Starting benchmark..."
+
+    # =============================================================================
+    # Run Benchmark
+    # =============================================================================
+    cd ../../../benchmarks/
+    python3 benchmark_serving.py --port 10001 --seed $(date +%s) \
+        --model $MODEL \
+        --dataset-name random --random-input-len 7500 --random-output-len 200 \
+        --num-prompts 200 --burstiness 100 --request-rate 2 | tee benchmark.log
+
+    echo "Benchmarking done. Cleaning up..."
+
+    cleanup
+}
+
+main

examples/online_serving/openai_transcription_client.py

@@ -19,10 +19,8 @@ The script performs:
 """
 
 import asyncio
-import json
 
-import httpx
-from openai import OpenAI
+from openai import AsyncOpenAI, OpenAI
 
 from vllm.assets.audio import AudioAsset
 
@@ -47,37 +45,30 @@ def sync_openai(audio_path: str, client: OpenAI):
         print("transcription result:", transcription.text)
 
 
-async def stream_openai_response(audio_path: str, base_url: str, api_key: str):
+async def stream_openai_response(audio_path: str, client: AsyncOpenAI):
     """
-    Perform streaming transcription using vLLM's raw HTTP streaming API.
+    Perform asynchronous transcription using OpenAI-compatible API.
     """
-    data = {
-        "language": "en",
-        "stream": True,
-        "model": "openai/whisper-large-v3",
-    }
-    url = base_url + "/audio/transcriptions"
-    headers = {"Authorization": f"Bearer {api_key}"}
-    print("transcription result:", end=" ")
-    # OpenAI Transcription API client does not support streaming.
-    async with httpx.AsyncClient() as client:
-        with open(audio_path, "rb") as f:
-            async with client.stream(
-                "POST", url, files={"file": f}, data=data, headers=headers
-            ) as response:
-                async for line in response.aiter_lines():
-                    # Each line is a JSON object prefixed with 'data: '
-                    if line:
-                        if line.startswith("data: "):
-                            line = line[len("data: ") :]
-                        # Last chunk, stream ends
-                        if line.strip() == "[DONE]":
-                            break
-                        # Parse the JSON response
-                        chunk = json.loads(line)
-                        # Extract and print the content
-                        content = chunk["choices"][0].get("delta", {}).get("content")
-                        print(content, end="")
+    print("\ntranscription result:", end=" ")
+    with open(audio_path, "rb") as f:
+        transcription = await client.audio.transcriptions.create(
+            file=f,
+            model="openai/whisper-large-v3",
+            language="en",
+            response_format="json",
+            temperature=0.0,
+            # Additional sampling params not provided by OpenAI API.
+            extra_body=dict(
+                seed=420,
+                top_p=0.6,
+            ),
+            stream=True,
+        )
+        async for chunk in transcription:
+            if chunk.choices:
+                content = chunk.choices[0].get("delta", {}).get("content")
+                print(content, end="", flush=True)
+
     print()  # Final newline after stream ends
 
 
@@ -95,7 +86,11 @@ def main():
 
     sync_openai(mary_had_lamb, client)
     # Run the asynchronous function
-    asyncio.run(stream_openai_response(winning_call, openai_api_base, openai_api_key))
+    client = AsyncOpenAI(
+        api_key=openai_api_key,
+        base_url=openai_api_base,
+    )
+    asyncio.run(stream_openai_response(winning_call, client))
 
 
 if __name__ == "__main__":

mkdocs.yaml

@@ -127,6 +127,7 @@ extra_javascript:
   - mkdocs/javascript/run_llm_widget.js
   - https://cdn.mathjax.org/mathjax/latest/MathJax.js?config=TeX-AMS_HTML
   - mkdocs/javascript/edit_and_feedback.js
+  - mkdocs/javascript/slack_and_forum.js
 
 # Makes the url format end in .html rather than act as a dir
 # So index.md generates as index.html and is available under URL /index.html

pyproject.toml

@@ -76,7 +76,7 @@ line-length = 80
 "vllm/spec_decode/**/*.py" = ["UP006", "UP035"]
 "vllm/worker/**/*.py" = ["UP006", "UP035"]
 # Python 3.8 typing - skip utils for ROCm
-"vllm/utils.py" = ["UP006", "UP035"]
+"vllm/utils/__init__.py" = ["UP006", "UP035"]
 
 [tool.ruff.lint]
 select = [
@@ -150,6 +150,7 @@ skip_gitignore = true
 markers = [
     "skip_global_cleanup",
     "core_model: enable this model test in each PR instead of only nightly",
+    "hybrid_model: models that contain mamba layers (including pure SSM and hybrid architectures)",
     "cpu_model: enable this model test in CPU tests",
     "split: run this test as part of a split",
     "distributed: run this test only in distributed GPU tests",

requirements/common.txt

@@ -23,7 +23,7 @@ lm-format-enforcer >= 0.10.11, < 0.11
 llguidance >= 0.7.11, < 0.8.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64"
 outlines == 0.1.11
 lark == 1.2.2
-xgrammar == 0.1.19; platform_machine == "x86_64" or platform_machine == "aarch64"
+xgrammar == 0.1.19; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64"
 typing_extensions >= 4.10
 filelock >= 3.16.1 # need to contain https://github.com/tox-dev/filelock/pull/317
 partial-json-parser # used for parsing partial JSON outputs

tests/async_engine/test_async_llm_engine.py

@@ -6,8 +6,8 @@ import os
 import uuid
 from asyncio import CancelledError
 from copy import copy
-from dataclasses import dataclass
-from typing import Optional
+from dataclasses import dataclass, field
+from typing import Any, Optional
 
 import pytest
 import pytest_asyncio
@@ -32,6 +32,8 @@ class RequestOutput:
 @dataclass
 class MockModelConfig:
     use_async_output_proc = True
+    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
+    mm_placeholder_str_override: dict[str, str] = field(default_factory=dict)
 
 
 class MockEngine:

tests/build_cython.py

@@ -25,7 +25,7 @@ infiles += [
 infiles += [
     "vllm/model_executor/layers/sampler.py",
     "vllm/sampling_params.py",
-    "vllm/utils.py",
+    "vllm/utils/__init__.py",
 ]
 
 setup(ext_modules=cythonize(infiles,

tests/conftest.py

@@ -34,7 +34,6 @@ from vllm.logger import init_logger
 from vllm.outputs import RequestOutput
 from vllm.sampling_params import BeamSearchParams
 from vllm.transformers_utils.utils import maybe_model_redirect
-from vllm.utils import cuda_device_count_stateless
 
 logger = init_logger(__name__)
 
@@ -1094,7 +1093,8 @@ def num_gpus_available():
     """Get number of GPUs without initializing the CUDA context
     in current process."""
 
-    return cuda_device_count_stateless()
+    from vllm.platforms import current_platform
+    return current_platform.device_count()
 
 
 temp_dir = tempfile.gettempdir()

tests/engine/test_arg_utils.py

@@ -231,6 +231,58 @@ def test_limit_mm_per_prompt_parser(arg, expected):
     assert args.limit_mm_per_prompt == expected
 
 
+@pytest.mark.parametrize(
+    ("arg", "expected"),
+    [
+        (None, dict()),
+        ('{"video": {"num_frames": 123} }', {
+            "video": {
+                "num_frames": 123
+            }
+        }),
+        (
+            '{"video": {"num_frames": 123, "fps": 1.0, "foo": "bar"}, "image": {"foo": "bar"} }',  # noqa
+            {
+                "video": {
+                    "num_frames": 123,
+                    "fps": 1.0,
+                    "foo": "bar"
+                },
+                "image": {
+                    "foo": "bar"
+                }
+            }),
+    ])
+def test_media_io_kwargs_parser(arg, expected):
+    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
+    if arg is None:
+        args = parser.parse_args([])
+    else:
+        args = parser.parse_args(["--media-io-kwargs", arg])
+
+    assert args.media_io_kwargs == expected
+
+
+@pytest.mark.parametrize(("arg", "expected"), [
+    (None, dict()),
+    ('{"video":"<|video_placeholder|>"}', {
+        "video": "<|video_placeholder|>"
+    }),
+    ('{"video":"<|video_placeholder|>", "image": "<|image_placeholder|>"}', {
+        "video": "<|video_placeholder|>",
+        "image": "<|image_placeholder|>"
+    }),
+])
+def test_mm_placeholder_str_override_parser(arg, expected):
+    parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
+    if arg is None:
+        args = parser.parse_args([])
+    else:
+        args = parser.parse_args(["--mm-placeholder-str-override", arg])
+
+    assert args.mm_placeholder_str_override == expected
+
+
 def test_compilation_config():
     parser = EngineArgs.add_cli_args(FlexibleArgumentParser())
 
@@ -239,32 +291,40 @@ def test_compilation_config():
     assert args.compilation_config == CompilationConfig()
 
     # set to O3
-    args = parser.parse_args(["-O3"])
-    assert args.compilation_config.level == 3
+    args = parser.parse_args(["-O0"])
+    assert args.compilation_config.level == 0
 
     # set to O 3 (space)
-    args = parser.parse_args(["-O", "3"])
-    assert args.compilation_config.level == 3
+    args = parser.parse_args(["-O", "1"])
+    assert args.compilation_config.level == 1
 
     # set to O 3 (equals)
-    args = parser.parse_args(["-O=3"])
+    args = parser.parse_args(["-O=2"])
+    assert args.compilation_config.level == 2
+
+    # set to O.level 3
+    args = parser.parse_args(["-O.level", "3"])
     assert args.compilation_config.level == 3
 
     # set to string form of a dict
     args = parser.parse_args([
-        "--compilation-config",
-        '{"level": 3, "cudagraph_capture_sizes": [1, 2, 4, 8]}',
+        "-O",
+        '{"level": 3, "cudagraph_capture_sizes": [1, 2, 4, 8], '
+        '"use_inductor": false}',
     ])
     assert (args.compilation_config.level == 3 and
-            args.compilation_config.cudagraph_capture_sizes == [1, 2, 4, 8])
+            args.compilation_config.cudagraph_capture_sizes == [1, 2, 4, 8]
+            and not args.compilation_config.use_inductor)
 
     # set to string form of a dict
     args = parser.parse_args([
         "--compilation-config="
-        '{"level": 3, "cudagraph_capture_sizes": [1, 2, 4, 8]}',
+        '{"level": 3, "cudagraph_capture_sizes": [1, 2, 4, 8], '
+        '"use_inductor": true}',
     ])
     assert (args.compilation_config.level == 3 and
-            args.compilation_config.cudagraph_capture_sizes == [1, 2, 4, 8])
+            args.compilation_config.cudagraph_capture_sizes == [1, 2, 4, 8]
+            and args.compilation_config.use_inductor)
 
 
 def test_prefix_cache_default():

tests/entrypoints/openai/test_completion.py

@@ -779,3 +779,57 @@ async def test_guided_decoding_type_error(client: openai.AsyncOpenAI,
             prompt="Give an example string that fits this regex",
             extra_body=dict(guided_regex=sample_regex,
                             guided_json=sample_json_schema))
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "model_name,stream,echo",
+    [
+        (MODEL_NAME, False, False),
+        (MODEL_NAME, False, True),
+        (MODEL_NAME, True, False),
+        (MODEL_NAME, True, True)  # should not raise BadRequestError error
+    ],
+)
+async def test_echo_stream_completion(client: openai.AsyncOpenAI,
+                                      model_name: str, stream: bool,
+                                      echo: bool):
+    saying: str = "Hello, my name is"
+    result = await client.completions.create(model=model_name,
+                                             prompt=saying,
+                                             max_tokens=10,
+                                             temperature=0.0,
+                                             echo=echo,
+                                             stream=stream)
+
+    stop_reason = "length"
+
+    if not stream:
+        completion = result
+        assert completion.id is not None
+        assert completion.choices is not None and len(completion.choices) == 1
+
+        choice = completion.choices[0]
+        assert len(choice.text) >= 5
+        assert choice.finish_reason == stop_reason
+
+        if echo:
+            assert choice.text is not None and saying in choice.text
+        else:
+            assert choice.text is not None and saying not in choice.text
+
+    else:
+        chunks: list[str] = []
+        final_finish_reason = None
+        async for chunk in result:
+            if chunk.choices and chunk.choices[0].text:
+                chunks.append(chunk.choices[0].text)
+            if chunk.choices and chunk.choices[0].finish_reason:
+                final_finish_reason = chunk.choices[0].finish_reason
+
+        assert final_finish_reason == stop_reason
+        content = "".join(chunks)
+        if echo:
+            assert content is not None and saying in content
+        else:
+            assert content is not None and saying not in content

tests/entrypoints/openai/test_serving_chat.py

@@ -3,8 +3,8 @@
 
 import asyncio
 from contextlib import suppress
-from dataclasses import dataclass
-from typing import Optional
+from dataclasses import dataclass, field
+from typing import Any, Optional
 from unittest.mock import MagicMock
 
 from vllm.config import MultiModalConfig
@@ -40,6 +40,8 @@ class MockModelConfig:
     allowed_local_media_path: str = ""
     encoder_config = None
     generation_config: str = "auto"
+    media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
+    mm_placeholder_str_override: dict[str, str] = field(default_factory=dict)
 
     def get_diff_sampling_param(self):
         return self.diff_sampling_param or {}

tests/entrypoints/openai/test_video.py

@@ -50,7 +50,7 @@ async def client(server):
 @pytest.fixture(scope="session")
 def base64_encoded_video() -> dict[str, str]:
     return {
-        video_url: encode_video_base64(fetch_video(video_url))
+        video_url: encode_video_base64(fetch_video(video_url)[0])
         for video_url in TEST_VIDEO_URLS
     }
 

tests/kernels/attention/test_attention_selector.py

@@ -106,10 +106,8 @@ def test_env(
                                                    block_size,
                                                    False,
                                                    use_mla=use_mla)
-                        if use_v1 and name != "TRITON_MLA":
-                            assert backend.get_name() == f"{name}_VLLM_V1"
-                        else:
-                            assert backend.get_name() == name
+                        expected = f"{name}_VLLM_V1" if use_v1 else name
+                        assert backend.get_name() == expected
                     else:
                         with pytest.raises(ValueError) as exc_info:
                             get_attn_backend(16,

tests/kernels/attention/test_rocm_attention_selector.py

@@ -35,7 +35,8 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
         m.setenv(STR_BACKEND_ENV_VAR, "TRITON_MLA")
         backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False,
                                    False, True)
-        assert backend.get_name() == "TRITON_MLA"
+        assert (backend.get_name() == "TRITON_MLA"
+                or backend.get_name() == "TRITON_MLA_VLLM_V1")
 
         # If attention backend is None
         # If use_mla is true
@@ -43,7 +44,8 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
         m.setenv(STR_BACKEND_ENV_VAR, None)
         backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False,
                                    False, True)
-        assert backend.get_name() == "TRITON_MLA"
+        assert (backend.get_name() == "TRITON_MLA"
+                or backend.get_name() == "TRITON_MLA_VLLM_V1")
 
         # change the attention backend to AITER MLA
         m.setenv(STR_BACKEND_ENV_VAR, "ROCM_AITER_MLA")

tests/kernels/moe/parallel_utils.py

@@ -0,0 +1,190 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+DeepEP test utilities
+"""
+import dataclasses
+import importlib
+import os
+import traceback
+from typing import Callable, Optional
+
+import torch
+from torch.distributed import ProcessGroup
+from torch.multiprocessing import (
+    spawn)  # pyright: ignore[reportPrivateImportUsage]
+from typing_extensions import Concatenate, ParamSpec
+
+from vllm.utils import get_open_port
+
+has_deep_ep = importlib.util.find_spec("deep_ep") is not None
+if has_deep_ep:
+    from vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize import (  # noqa: E501
+        DeepEPHTPrepareAndFinalize)
+    from vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize import (  # noqa: E501
+        DeepEPLLPrepareAndFinalize)
+
+## Parallel Processes Utils
+
+P = ParamSpec("P")
+
+
+@dataclasses.dataclass
+class ProcessGroupInfo:
+    world_size: int
+    world_local_size: int
+    rank: int
+    node_rank: int
+    local_rank: int
+    device: torch.device
+
+
+def _worker_parallel_launch(
+    local_rank: int,
+    world_size: int,
+    world_local_size: int,
+    node_rank: int,
+    init_method: str,
+    worker: Callable[Concatenate[ProcessGroupInfo, P], None],
+    *args: P.args,
+    **kwargs: P.kwargs,
+) -> None:
+    rank = node_rank * world_local_size + local_rank
+    torch.cuda.set_device(local_rank)
+    device = torch.device("cuda", local_rank)
+    torch.distributed.init_process_group(
+        backend="cpu:gloo,cuda:nccl",
+        init_method=init_method,
+        rank=rank,
+        world_size=world_size,
+        device_id=device,
+    )
+    barrier = torch.tensor([rank], device=device)
+    torch.distributed.all_reduce(barrier)
+
+    try:
+        worker(
+            ProcessGroupInfo(
+                world_size=world_size,
+                world_local_size=world_local_size,
+                rank=rank,
+                node_rank=node_rank,
+                local_rank=local_rank,
+                device=device,
+            ),
+            *args,
+            **kwargs,
+        )
+    except Exception as ex:
+        print(ex)
+        traceback.print_exc()
+        raise
+    finally:
+        torch.distributed.destroy_process_group()
+
+
+def parallel_launch(
+    world_size: int,
+    worker: Callable[Concatenate[ProcessGroupInfo, P], None],
+    *args: P.args,
+    **kwargs: P.kwargs,
+) -> None:
+    assert not kwargs
+    spawn(
+        _worker_parallel_launch,
+        args=(
+            world_size,
+            world_size,
+            0,
+            f"tcp://{os.getenv('LOCALHOST', 'localhost')}:{get_open_port()}",
+            worker,
+        ) + args,
+        nprocs=world_size,
+        join=True,
+    )
+
+
+## DeepEP specific utils
+
+
+@dataclasses.dataclass
+class DeepEPHTArgs:
+    num_local_experts: int
+
+
+@dataclasses.dataclass
+class DeepEPLLArgs:
+    max_tokens_per_rank: int
+    hidden_size: int
+    num_experts: int
+    use_fp8_dispatch: bool
+
+
+def make_deepep_ht_a2a(pg: ProcessGroup,
+                       pgi: ProcessGroupInfo,
+                       dp_size: int,
+                       ht_args: DeepEPHTArgs,
+                       q_dtype: Optional[torch.dtype] = None,
+                       block_shape: Optional[list[int]] = None):
+
+    import deep_ep
+
+    # high throughput a2a
+    num_nvl_bytes = 1024 * 1024 * 1024  # 1GB
+    num_rdma_bytes, low_latency_mode, num_qps_per_rank = 0, False, 1
+    buffer = deep_ep.Buffer(group=pg,
+                            num_nvl_bytes=num_nvl_bytes,
+                            num_rdma_bytes=num_rdma_bytes,
+                            low_latency_mode=low_latency_mode,
+                            num_qps_per_rank=num_qps_per_rank)
+    return DeepEPHTPrepareAndFinalize(buffer=buffer,
+                                      world_size=pgi.world_size,
+                                      rank=pgi.rank,
+                                      dp_size=dp_size,
+                                      rank_expert_offset=pgi.rank *
+                                      ht_args.num_local_experts)
+
+
+def make_deepep_ll_a2a(pg: ProcessGroup,
+                       pgi: ProcessGroupInfo,
+                       dp_size: int,
+                       deepep_ll_args: DeepEPLLArgs,
+                       q_dtype: Optional[torch.dtype] = None,
+                       block_shape: Optional[list[int]] = None):
+
+    import deep_ep
+
+    # low-latency a2a
+    num_rdma_bytes = deep_ep.Buffer.get_low_latency_rdma_size_hint(
+        deepep_ll_args.max_tokens_per_rank, deepep_ll_args.hidden_size,
+        pgi.world_size, deepep_ll_args.num_experts)
+
+    buffer = deep_ep.Buffer(group=pg,
+                            num_rdma_bytes=num_rdma_bytes,
+                            low_latency_mode=True,
+                            num_qps_per_rank=deepep_ll_args.num_experts //
+                            pgi.world_size)
+
+    return DeepEPLLPrepareAndFinalize(
+        buffer=buffer,
+        world_size=pgi.world_size,
+        dp_size=dp_size,
+        max_tokens_per_rank=deepep_ll_args.max_tokens_per_rank,
+        use_fp8_dispatch=deepep_ll_args.use_fp8_dispatch,
+    )
+
+
+def make_deepep_a2a(pg: ProcessGroup,
+                    pgi: ProcessGroupInfo,
+                    dp_size: int,
+                    deepep_ht_args: Optional[DeepEPHTArgs],
+                    deepep_ll_args: Optional[DeepEPLLArgs],
+                    q_dtype: Optional[torch.dtype] = None,
+                    block_shape: Optional[list[int]] = None):
+    if deepep_ht_args is not None:
+        assert deepep_ll_args is None
+        return make_deepep_ht_a2a(pg, pgi, dp_size, deepep_ht_args, q_dtype,
+                                  block_shape)
+
+    assert deepep_ll_args is not None
+    return make_deepep_ll_a2a(pg, pgi, dp_size, deepep_ll_args, q_dtype,
+                              block_shape)

tests/kernels/moe/test_batched_moe.py

@@ -2,18 +2,59 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from dataclasses import dataclass
+from typing import Optional
 
 import pytest
 import torch
 import triton.language as tl
 
+from tests.kernels.moe.utils import (batched_moe,
+                                     make_quantized_test_activations,
+                                     make_test_weights, triton_moe)
+from tests.kernels.quant_utils import native_batched_masked_quant_matmul
+from tests.kernels.utils import torch_experts
+from vllm.config import VllmConfig, set_current_vllm_config
 from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
     invoke_moe_batched_triton_kernel)
+from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
+from vllm.platforms import current_platform
+
+MNK_FACTORS = [
+    (1, 128, 128),
+    (1, 128, 2048),
+    (1, 512, 512),
+    (1, 1024, 128),
+    (1, 1024, 2048),
+    (32, 128, 128),
+    (32, 512, 512),
+    (32, 1024, 2048),
+    (45, 128, 128),
+    (45, 128, 2048),
+    (45, 512, 512),
+    (45, 1024, 128),
+    (45, 1024, 2048),
+    (64, 128, 128),
+    (64, 512, 512),
+    (64, 1024, 2048),
+    (222, 128, 128),
+    (222, 128, 2048),
+    (222, 512, 512),
+    (222, 1024, 128),
+    (222, 1024, 2048),
+]
+NUM_EXPERTS = [8, 64]
+TOP_KS = [1, 2, 6]
+
+vllm_config = VllmConfig()
+vllm_config.scheduler_config.max_num_seqs = 128
+vllm_config.scheduler_config.max_model_len = 8192
 
 
 @dataclass
 class BatchedMMConfig:
-    dtype: torch.dtype
+    in_dtype: torch.dtype
+    quant_dtype: Optional[torch.dtype]
+    out_dtype: torch.dtype
     num_experts: int
     max_tokens_per_expert: int
     K: int
@@ -32,79 +73,127 @@ class BatchedMMTensors:
         A = torch.randn(
             (config.num_experts, config.max_tokens_per_expert, config.K),
             device="cuda",
-            dtype=config.dtype) / 10
+            dtype=config.in_dtype) / 10
         B = torch.randn((config.num_experts, config.N, config.K),
                         device="cuda",
-                        dtype=config.dtype)
+                        dtype=config.in_dtype)
         C = torch.zeros(
             (config.num_experts, config.max_tokens_per_expert, config.N),
             device="cuda",
-            dtype=config.dtype)
+            dtype=config.out_dtype)
+
         num_expert_tokens = torch.randint(low=0,
                                           high=config.max_tokens_per_expert,
                                           size=(config.num_experts, ),
                                           device="cuda",
                                           dtype=torch.int32)
+
         return BatchedMMTensors(A, B, C, num_expert_tokens)
 
 
-def ref_impl(A: torch.Tensor, B: torch.Tensor, C: torch.Tensor,
-             num_expert_tokens: torch.Tensor) -> torch.Tensor:
-
-    num_expert_tokens_cpu = num_expert_tokens.clone()
-    num_expert_tokens_cpu = num_expert_tokens_cpu.to(device="cpu")
-    num_experts = num_expert_tokens.size(0)
-
-    for e in range(num_experts):
-        num_tokens = num_expert_tokens_cpu[e]
-        C[e, :num_tokens, :] = A[e, :num_tokens, :] @ B[e].transpose(0, 1)
-
-    return C
-
-
-@pytest.mark.parametrize("num_experts", [16, 32])
+@pytest.mark.parametrize("num_experts", [8, 16, 32])
 @pytest.mark.parametrize("max_tokens_per_expert",
                          [32, 64, 128, 192, 224, 256, 512])
 @pytest.mark.parametrize("K", [128, 256, 1024])
 @pytest.mark.parametrize("N", [128, 256, 512, 1024])
 @pytest.mark.parametrize("dtype",
                          [torch.float32, torch.float16, torch.bfloat16])
+@pytest.mark.parametrize("block_shape", [None])
+@pytest.mark.parametrize("per_act_token_quant", [False])
 def test_batched_mm(num_experts: int, max_tokens_per_expert: int, K: int,
-                    N: int, dtype: torch.dtype):
+                    N: int, dtype: torch.dtype,
+                    block_shape: Optional[list[int]],
+                    per_act_token_quant: bool):
+    current_platform.seed_everything(7)
 
-    config = BatchedMMConfig(dtype, num_experts, max_tokens_per_expert, K, N)
-    tensors = BatchedMMTensors.make_tensors(config)
+    use_fp8_w8a8 = dtype == torch.float8_e4m3fn
 
-    test_output = tensors.C
-    ref_output = test_output.clone()
+    if (per_act_token_quant or block_shape is not None) and not use_fp8_w8a8:
+        pytest.skip("Don't test blocking for non-quantized types.")
+
+    if per_act_token_quant and block_shape is not None:
+        pytest.skip("Skip illegal quantization test.")
+
+    if dtype.itemsize == 1:
+        act_dtype = torch.bfloat16
+        quant_dtype = dtype
+    else:
+        act_dtype = dtype
+        quant_dtype = None
+
+    num_expert_tokens = torch.randint(low=0,
+                                      high=max_tokens_per_expert,
+                                      size=(num_experts, ),
+                                      device="cuda",
+                                      dtype=torch.int32)
+
+    A, A_q, A_scale = make_quantized_test_activations(
+        num_experts,
+        max_tokens_per_expert,
+        K,
+        in_dtype=act_dtype,
+        quant_dtype=quant_dtype,
+        block_shape=block_shape,
+        per_act_token_quant=per_act_token_quant)
+
+    B, B_q, B_scale, _, _, _ = make_test_weights(
+        num_experts,
+        N // 2,
+        K,
+        in_dtype=act_dtype,
+        quant_dtype=quant_dtype,
+        block_shape=block_shape,
+    )
+
+    out_shape = (num_experts, max_tokens_per_expert, N)
+    test_output = torch.zeros(out_shape, dtype=act_dtype, device="cuda")
+    ref_output = torch.zeros(out_shape, dtype=act_dtype, device="cuda")
+    q_ref_output = torch.zeros(out_shape, dtype=act_dtype, device="cuda")
 
     compute_tl_dtype = {
         torch.float16: tl.float16,
         torch.bfloat16: tl.bfloat16,
         torch.float32: tl.float32
     }[test_output.dtype]
+
+    assert A_q.dtype == B_q.dtype
+
     invoke_moe_batched_triton_kernel(
-        tensors.A,
-        tensors.B,
+        A_q,
+        B_q,
         test_output,
-        tensors.num_expert_tokens,
+        num_expert_tokens,
         compute_tl_dtype,
         # Quantization data
-        None,
-        None,
+        A_scale,
+        B_scale,
         None,
         # Quantization schemes
-        False,
+        use_fp8_w8a8,
         False,
         False,
         config={
             "BLOCK_SIZE_M": 16,
             "BLOCK_SIZE_N": 16,
-            "BLOCK_SIZE_K": 16
-        })
+            "BLOCK_SIZE_K": 16 if dtype.itemsize > 1 else 32
+        },
+        block_shape=block_shape,
+    )
 
-    ref_output = ref_impl(tensors.A, tensors.B, ref_output,
-                          tensors.num_expert_tokens)
+    ref_output = native_batched_masked_quant_matmul(
+        A,
+        B,
+        ref_output,
+        num_expert_tokens,
+        None,
+        None,
+        None,
+    )
+
+    q_ref_output = native_batched_masked_quant_matmul(A_q, B_q, q_ref_output,
+                                                      num_expert_tokens,
+                                                      A_scale, B_scale,
+                                                      block_shape)
 
     rtol, atol = {
         torch.float16: (6e-2, 6e-2),
@@ -112,4 +201,98 @@ def test_batched_mm(num_experts: int, max_tokens_per_expert: int, K: int,
         torch.float32: (1e-2, 1e-2),
     }[test_output.dtype]
 
-    torch.testing.assert_close(test_output, ref_output, atol=atol, rtol=rtol)
+    torch.testing.assert_close(ref_output, test_output, atol=atol, rtol=rtol)
+    torch.testing.assert_close(test_output, q_ref_output, atol=atol, rtol=rtol)
+
+
+@pytest.mark.parametrize(("m", "n", "k"), MNK_FACTORS)
+@pytest.mark.parametrize("e", NUM_EXPERTS)
+@pytest.mark.parametrize("topk", TOP_KS)
+@pytest.mark.parametrize("dtype", [torch.bfloat16])
+@pytest.mark.parametrize("per_act_token_quant", [False])
+@pytest.mark.parametrize("block_shape", [None])
+def test_fused_moe_batched_experts(
+    m: int,
+    n: int,
+    k: int,
+    e: int,
+    topk: int,
+    dtype: torch.dtype,
+    per_act_token_quant: bool,
+    block_shape: Optional[list[int]],
+):
+    current_platform.seed_everything(7)
+
+    use_fp8_w8a8 = dtype == torch.float8_e4m3fn
+
+    if not use_fp8_w8a8 and (per_act_token_quant or block_shape is not None):
+        pytest.skip("Skip quantization test for non-quantized type")
+
+    if per_act_token_quant and block_shape is not None or topk > e:
+        pytest.skip("Skip illegal quantization test.")
+
+    a = torch.randn((m, k), device="cuda", dtype=torch.bfloat16) / 10
+    score = torch.randn((m, e), device="cuda", dtype=torch.bfloat16)
+
+    if dtype.itemsize == 1:
+        act_dtype = torch.bfloat16
+        quant_dtype = dtype
+    else:
+        act_dtype = dtype
+        quant_dtype = None
+
+    _, w1, w1_s, _, w2, w2_s = make_test_weights(e,
+                                                 n,
+                                                 k,
+                                                 block_shape=block_shape,
+                                                 in_dtype=act_dtype,
+                                                 quant_dtype=quant_dtype)
+
+    with set_current_vllm_config(vllm_config):
+        topk_weight, topk_ids, _ = fused_topk(a, score, topk, False)
+        batched_output = batched_moe(
+            a,
+            w1,
+            w2,
+            topk_weight,
+            topk_ids,
+            w1_scale=w1_s,
+            w2_scale=w2_s,
+            quant_dtype=quant_dtype,
+            per_act_token_quant=per_act_token_quant,
+            block_shape=block_shape,
+        )
+        baseline_output = torch_experts(
+            a,
+            w1,
+            w2,
+            topk_weight,
+            topk_ids,
+            w1_scale=w1_s,
+            w2_scale=w2_s,
+            quant_dtype=quant_dtype,
+            per_act_token_quant=per_act_token_quant,
+            block_shape=block_shape)
+
+        triton_output = triton_moe(
+            a,
+            w1,
+            w2,
+            topk_weight,
+            topk_ids,
+            w1_scale=w1_s,
+            w2_scale=w2_s,
+            quant_dtype=quant_dtype,
+            per_act_token_quant=per_act_token_quant,
+            block_shape=block_shape,
+        )
+
+    torch.testing.assert_close(triton_output,
+                               baseline_output,
+                               atol=2e-2,
+                               rtol=2e-2)
+
+    torch.testing.assert_close(triton_output,
+                               batched_output,
+                               atol=2e-2,
+                               rtol=2e-2)

tests/kernels/moe/test_block_fp8.py

@@ -0,0 +1,296 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+import torch
+
+from tests.kernels.moe.utils import make_test_weights
+from tests.kernels.quant_utils import (native_per_token_group_quant_fp8,
+                                       native_w8a8_block_matmul)
+from vllm.config import VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.fused_moe import fused_experts
+from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
+    _valid_deep_gemm_shape, deep_gemm_moe_fp8)
+from vllm.model_executor.layers.fused_moe.fused_moe import (
+    fused_topk, modular_triton_fused_moe)
+from vllm.platforms import current_platform
+
+dg_available = False
+try:
+    import deep_gemm
+    dg_available = True
+except ImportError:
+    pass
+
+if current_platform.get_device_capability() < (9, 0):
+    pytest.skip("FP8 Triton requires CUDA 9.0 or higher",
+                allow_module_level=True)
+
+vllm_config = VllmConfig()
+vllm_config.scheduler_config.max_num_seqs = 128
+vllm_config.scheduler_config.max_model_len = 8192
+
+# Test configurations
+DTYPES = [torch.bfloat16]  # [torch.half, torch.bfloat16, torch.float32]
+# Deepseek-V3's intermediate size 18432, so N is 18432*2/8=4608 at TP8
+# and its hidden size is 7168.
+MNK_FACTORS = [
+    (1, 128, 128),
+    (1, 512, 512),
+    (1, 128, 7168),
+    (1, 1024, 7168),
+    (1, 4608, 128),
+    (1, 4608, 512),
+    (1, 4608, 7168),
+    (83, 128, 128),
+    (83, 512, 512),
+    (83, 1024, 7168),
+    (83, 4608, 512),
+    (83, 4608, 7168),
+    (128, 128, 128),
+    (128, 512, 512),
+    (128, 1024, 7168),
+    (128, 4608, 512),
+    (128, 4608, 7168),
+    (2048, 128, 128),
+    (2048, 1024, 7168),
+    (2048, 4608, 512),
+    (2048, 4608, 7168),
+    (8192, 128, 128),
+    (8192, 512, 512),
+    (8192, 128, 7168),
+    (8192, 1024, 7168),
+    (8192, 4608, 512),
+    (8192, 4608, 7168),
+]
+
+MNK_FACTORS_DG = [
+    (128, 128, 128),
+    (128, 512, 512),
+    (128, 128, 7168),
+    (128, 1024, 7168),
+    (128, 4608, 128),
+    (128, 4608, 512),
+    (128, 4608, 7168),
+    (192, 128, 128),
+    (192, 512, 512),
+    (192, 1024, 7168),
+    (192, 4608, 512),
+    (192, 4608, 7168),
+    (1335, 128, 128),
+    (1335, 1024, 7168),
+    (1335, 4608, 512),
+    (1335, 4608, 7168),
+    (2048, 128, 128),
+    (2048, 512, 512),
+    (2048, 128, 7168),
+    (2048, 1024, 7168),
+    (2048, 4608, 128),
+    (2048, 4608, 512),
+    (2048, 4608, 7168),
+]
+
+BLOCK_SIZE = [[128, 128]]
+E = [2, 8, 16]  # [128, 256]
+TOP_KS = [1, 2, 6]
+SEEDS = [0]
+
+
+def torch_w8a8_block_fp8_moe(a, w1, w2, w1_s, w2_s, topk_weight, topk_ids,
+                             block_shape):
+    """Fused moe with block-wise quantization using native torch."""
+    B, D = a.shape
+    topk = topk_ids.size(1)
+    a = a.view(B, -1, D).repeat(1, topk, 1).reshape(-1, D)
+    out = torch.zeros(B * topk, w2.shape[1], dtype=a.dtype, device=a.device)
+
+    topk_weight = topk_weight.view(-1)
+    topk_ids = topk_ids.view(-1)
+
+    _, block_k = block_shape[0], block_shape[1]
+    a_q, a_s = native_per_token_group_quant_fp8(a, block_k)
+    a_q = a_q.to(torch.float32)
+    for i in range(w1.shape[0]):
+        mask = topk_ids == i
+        if mask.sum():
+            inter_out = native_w8a8_block_matmul(a_q[mask],
+                                                 w1[i],
+                                                 a_s[mask],
+                                                 w1_s[i],
+                                                 block_shape,
+                                                 output_dtype=a.dtype)
+            act_out = SiluAndMul().forward_native(inter_out)
+            act_out_q, act_out_s = native_per_token_group_quant_fp8(
+                act_out, block_k)
+            out[mask] = native_w8a8_block_matmul(act_out_q,
+                                                 w2[i],
+                                                 act_out_s,
+                                                 w2_s[i],
+                                                 block_shape,
+                                                 output_dtype=a.dtype)
+    return (out.view(B, -1, w2.shape[1]) *
+            topk_weight.view(B, -1, 1).to(out.dtype)).sum(dim=1)
+
+
+# Skip all tests if CUDA is not available
+pytest.importorskip("torch.cuda")
+
+
+@pytest.fixture(autouse=True)
+def setup_cuda():
+    torch.set_default_device("cuda")
+
+
+@pytest.mark.parametrize(("M", "N", "K"), MNK_FACTORS)
+@pytest.mark.parametrize("E", E)
+@pytest.mark.parametrize("topk", TOP_KS)
+@pytest.mark.parametrize("block_size", BLOCK_SIZE)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_w8a8_block_fp8_fused_moe(M, N, K, E, topk, block_size, dtype, seed,
+                                  monkeypatch):
+    if topk > E:
+        pytest.skip(f"Skipping test; topk={topk} > E={E}")
+
+    torch.manual_seed(seed)
+
+    monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", "2048")
+
+    a = torch.randn((M, K), dtype=dtype) / 10
+    score = torch.randn((M, E), dtype=dtype)
+
+    _, w1, w1_s, _, w2, w2_s = make_test_weights(E,
+                                                 N,
+                                                 K,
+                                                 dtype,
+                                                 torch.float8_e4m3fn,
+                                                 per_act_token_quant=False,
+                                                 block_shape=block_size)
+
+    m_fused_moe = modular_triton_fused_moe(use_fp8_w8a8=True,
+                                           use_int8_w8a8=False,
+                                           use_int8_w8a16=False,
+                                           use_int4_w4a16=False,
+                                           per_act_token_quant=False,
+                                           block_shape=block_size)
+
+    topk_weights, topk_ids, _ = fused_topk(a, score.float(), topk, False)
+
+    # Set the context to avoid lots of warning spam.
+    with set_current_vllm_config(vllm_config):
+        ref_out = torch_w8a8_block_fp8_moe(
+            a,
+            w1,
+            w2,
+            w1_s,
+            w2_s,
+            topk_weights,
+            topk_ids,
+            block_size,
+        )
+
+        out = fused_experts(
+            a,
+            w1,
+            w2,
+            topk_weights,
+            topk_ids,
+            use_fp8_w8a8=True,
+            w1_scale=w1_s,
+            w2_scale=w2_s,
+            block_shape=block_size,
+        )
+
+        m_out = m_fused_moe(
+            a,
+            w1,
+            w2,
+            topk_weights,
+            topk_ids,
+            w1_scale=w1_s,
+            w2_scale=w2_s,
+        )
+
+    # 0.039 only needed for [40000-4608-7168-2-1-block_size852-dtype852-0]
+    tol = 0.035 if M < 40000 else 0.039
+    torch.testing.assert_close(out, ref_out, atol=tol, rtol=tol)
+    torch.testing.assert_close(m_out, ref_out, atol=tol, rtol=tol)
+
+
+@pytest.mark.parametrize(("M", "N", "K"), MNK_FACTORS_DG)
+@pytest.mark.parametrize("E", E)
+@pytest.mark.parametrize("topk", TOP_KS)
+@pytest.mark.parametrize("seed", SEEDS)
+@pytest.mark.skipif(not dg_available, reason="DeepGemm kernels not available.")
+@torch.inference_mode()
+def test_w8a8_block_fp8_deep_gemm_fused_moe(M, N, K, E, topk, seed,
+                                            monkeypatch):
+    if topk > E:
+        pytest.skip(f"Skipping test: topk={topk} > E={E}")
+
+    if not _valid_deep_gemm_shape(M, N, K):
+        pytest.skip(f"Skipping test: invalid size m={M}, n={N}, k={K}")
+
+    chunk_size = 1024
+
+    torch.manual_seed(seed)
+
+    monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", str(chunk_size))
+
+    block_m = deep_gemm.get_m_alignment_for_contiguous_layout()
+    block_size = [block_m, block_m]
+    dtype = torch.bfloat16
+
+    a = torch.randn((M, K), dtype=dtype) / 10
+    score = torch.randn((M, E), dtype=dtype)
+
+    _, w1, w1_s, _, w2, w2_s = make_test_weights(E,
+                                                 N,
+                                                 K,
+                                                 dtype,
+                                                 torch.float8_e4m3fn,
+                                                 per_act_token_quant=False,
+                                                 block_shape=block_size)
+
+    # Note: for now use_compile will error out if the problem size is
+    # large enough to trigger chunking. I'm leaving the flag and
+    # setup code in case we are able to revisit this later.
+    use_compile = False
+
+    use_cudagraph = (chunk_size < M and N >= 1024 and K >= 1024
+                     and current_platform.is_cuda_alike())
+
+    topk_weights, topk_ids, _ = fused_topk(a, score.float(), topk, False)
+
+    # Set the context to avoid lots of warning spam.
+    with set_current_vllm_config(vllm_config):
+        ref_out = torch_w8a8_block_fp8_moe(a, w1, w2, w1_s, w2_s, topk_weights,
+                                           topk_ids, block_size)
+
+        if use_compile:
+            deep_gemm_moe_fp8_fn = torch.compile(deep_gemm_moe_fp8,
+                                                 backend="inductor",
+                                                 fullgraph=True)
+            torch._dynamo.mark_dynamic(a, 0)
+            torch._dynamo.mark_dynamic(topk_weights, 0)
+            torch._dynamo.mark_dynamic(topk_ids, 0)
+        else:
+            deep_gemm_moe_fp8_fn = deep_gemm_moe_fp8
+
+        out = deep_gemm_moe_fp8_fn(a, w1, w2, w1_s, w2_s, topk_weights,
+                                   topk_ids)
+
+        if use_cudagraph:
+            out.fill_(0)
+            stream = torch.cuda.Stream()
+            graph = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(graph, stream=stream):
+                out = deep_gemm_moe_fp8_fn(a, w1, w2, w1_s, w2_s, topk_weights,
+                                           topk_ids)
+            torch.cuda.synchronize()
+            graph.replay()
+            torch.cuda.synchronize()
+
+    torch.testing.assert_close(out, ref_out, atol=0.035, rtol=0.035)

tests/kernels/moe/test_block_int8.py

@@ -0,0 +1,147 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+import torch
+
+from tests.kernels.moe.utils import make_test_weights
+from tests.kernels.quant_utils import (native_per_token_group_quant_int8,
+                                       native_w8a8_block_matmul)
+from vllm.config import VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.fused_moe import fused_moe
+from vllm.platforms import current_platform
+
+if current_platform.get_device_capability() < (7, 0):
+    pytest.skip("INT8 Triton requires CUDA 7.0 or higher",
+                allow_module_level=True)
+
+vllm_config = VllmConfig()
+vllm_config.scheduler_config.max_num_seqs = 128
+vllm_config.scheduler_config.max_model_len = 8192
+
+DTYPES = [torch.half, torch.bfloat16]
+
+MNK_FACTORS = [
+    (1, 128, 128),
+    (1, 512, 512),
+    (1, 128, 7168),
+    (1, 1024, 7168),
+    (1, 4096, 128),
+    (1, 4096, 512),
+    (1, 4096, 7168),
+    (33, 128, 128),
+    (33, 512, 512),
+    (33, 128, 7168),
+    (33, 1024, 7168),
+    (33, 4096, 128),
+    (33, 4096, 512),
+    (33, 4096, 7168),
+    (128, 128, 128),
+    (128, 512, 512),
+    (128, 1024, 7168),
+    (128, 4096, 512),
+    (128, 4096, 7168),
+    (222, 128, 128),
+    (222, 512, 512),
+    (222, 1024, 7168),
+    (222, 4096, 512),
+    (222, 4096, 7168),
+    (2048, 128, 128),
+    (2048, 1024, 7168),
+    (2048, 4096, 512),
+    (2048, 4096, 7168),
+]
+
+E = [8, 24]
+TOP_KS = [2, 6]
+# BLOCK_SIZE = [[64, 64], [64, 128], [128, 64], [128, 128]]
+BLOCK_SIZE = [[128, 128]]
+SEEDS = [0]
+
+
+# For test
+def torch_w8a8_block_int8_moe(a, w1, w2, w1_s, w2_s, score, topk, block_shape):
+    """This function performs fused moe with block-wise quantization using
+    native torch."""
+    B, D = a.shape
+    a = a.view(B, -1, D).repeat(1, topk, 1).reshape(-1, D)
+    out = torch.zeros(B * topk, w2.shape[1], dtype=a.dtype, device=a.device)
+    score = torch.softmax(score, dim=-1, dtype=torch.float32)
+    topk_weight, topk_ids = torch.topk(score, topk)
+    topk_weight = topk_weight.view(-1)
+    topk_ids = topk_ids.view(-1)
+
+    _, block_k = block_shape[0], block_shape[1]
+    a_q, a_s = native_per_token_group_quant_int8(a, block_k)
+    for i in range(w1.shape[0]):
+        mask = topk_ids == i
+        if mask.sum():
+            inter_out = native_w8a8_block_matmul(a_q[mask],
+                                                 w1[i],
+                                                 a_s[mask],
+                                                 w1_s[i],
+                                                 block_shape,
+                                                 output_dtype=a.dtype)
+            act_out = SiluAndMul().forward_native(inter_out)
+            act_out_q, act_out_s = native_per_token_group_quant_int8(
+                act_out, block_k)
+            act_out = act_out.to(torch.float32)
+            out[mask] = native_w8a8_block_matmul(act_out_q,
+                                                 w2[i],
+                                                 act_out_s,
+                                                 w2_s[i],
+                                                 block_shape,
+                                                 output_dtype=a.dtype)
+    return (out.view(B, -1, w2.shape[1]) *
+            topk_weight.view(B, -1, 1).to(out.dtype)).sum(dim=1)
+
+
+@pytest.fixture(autouse=True, scope="module")
+def setup_cuda():
+    """Sets the default CUDA device for all tests in this module."""
+    torch.set_default_device("cuda")
+
+
+@pytest.mark.parametrize(("M", "N", "K"), MNK_FACTORS)
+@pytest.mark.parametrize("E", E)
+@pytest.mark.parametrize("topk", TOP_KS)
+@pytest.mark.parametrize("block_size", BLOCK_SIZE)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@torch.inference_mode()
+def test_w8a8_block_int8_fused_moe(M, N, K, E, topk, block_size, dtype, seed):
+    """Tests the fused_moe kernel with W8A8 INT8 block quantization against a
+    native torch reference."""
+    torch.manual_seed(seed)
+
+    a = torch.randn((M, K), dtype=dtype) / 10
+    score = torch.randn((M, E), dtype=dtype)
+
+    _, w1, w1_s, _, w2, w2_s = make_test_weights(E,
+                                                 N,
+                                                 K,
+                                                 dtype,
+                                                 torch.int8,
+                                                 per_act_token_quant=False,
+                                                 block_shape=block_size)
+
+    # Set the context to avoid lots of warning spam.
+    with set_current_vllm_config(vllm_config):
+        out = fused_moe(
+            a,
+            w1,
+            w2,
+            score,
+            topk,
+            renormalize=False,
+            use_int8_w8a8=True,
+            w1_scale=w1_s,
+            w2_scale=w2_s,
+            block_shape=block_size,
+        )
+        ref_out = torch_w8a8_block_int8_moe(a, w1, w2, w1_s, w2_s, score, topk,
+                                            block_size)
+
+    # Check results
+    torch.testing.assert_close(out, ref_out, atol=0.065, rtol=0.065)

tests/kernels/moe/test_cutlass_moe.py

@@ -97,11 +97,9 @@ class MOETensors8Bit(MOETensors):
         n_b_scales = 2 * n if per_out_channel else 1
         k_b_scales = k if per_out_channel else 1
         # Get the right scale for tests.
-        _, a_scale = ops.scaled_fp8_quant(
-            moe_tensors_fp16.a, use_per_token_if_dynamic=per_act_token)
-        a_q, _ = ops.scaled_fp8_quant(moe_tensors_fp16.a,
-                                      a_scale,
-                                      use_per_token_if_dynamic=per_act_token)
+        a_q, a_scale = ops.scaled_fp8_quant(
+            moe_tensors_fp16.a, None, use_per_token_if_dynamic=per_act_token)
+
         w1_q = torch.empty((e, 2 * n, k), device="cuda", dtype=q_dtype)
         w2_q = torch.empty((e, k, n), device="cuda", dtype=q_dtype)
 
@@ -187,6 +185,7 @@ def run_with_expert_maps(num_experts: int, num_local_experts: int,
 def run_8_bit(moe_tensors: MOETensors8Bit,
               topk_weights: torch.Tensor,
               topk_ids: torch.Tensor,
+              per_act_token: bool,
               num_local_experts: Optional[int] = None) -> torch.Tensor:
     assert not any([
         t is None for t in [
@@ -203,7 +202,8 @@ def run_8_bit(moe_tensors: MOETensors8Bit,
         'topk_ids': topk_ids,
         'w1_scale': moe_tensors.w1_scale,
         'w2_scale': moe_tensors.w2_scale,
-        'a1_scale': moe_tensors.a_scale
+        'per_act_token': per_act_token,
+        'a1_scale': None  #moe_tensors.a_scale
     }
 
     num_experts = moe_tensors.w1.size(0)
@@ -254,11 +254,13 @@ def test_cutlass_moe_8_bit_no_graph(
         triton_output = fused_experts(mt.a_d, mt.w1_d, mt.w2_d, topk_weights,
                                       topk_ids)
 
-        cutlass_output = run_8_bit(mt, topk_weights, topk_ids)
+        cutlass_output = run_8_bit(mt, topk_weights, topk_ids, per_act_token)
 
+        # Note 5.5 only needed for larger problem sizes, 5 works ok for
+        # the rest.
         torch.testing.assert_close(triton_output,
                                    cutlass_output,
-                                   atol=5e-2,
+                                   atol=5.5e-2,
                                    rtol=1e-2)
 
 
@@ -303,7 +305,8 @@ def test_cutlass_moe_8_bit_cuda_graph(
         stream = torch.cuda.Stream()
         graph = torch.cuda.CUDAGraph()
         with torch.cuda.graph(graph, stream=stream):
-            cutlass_output = run_8_bit(mt, topk_weights, topk_ids)
+            cutlass_output = run_8_bit(mt, topk_weights, topk_ids,
+                                       per_act_token)
 
         torch.cuda.synchronize()
         graph.replay()
@@ -359,6 +362,7 @@ def test_cutlass_moe_8_bit_EP(
         cutlass_output = run_8_bit(mt,
                                    topk_weights,
                                    topk_ids,
+                                   per_act_token,
                                    num_local_experts=e // ep_size)
 
         torch.testing.assert_close(triton_output,

tests/kernels/moe/test_deepep_deepgemm_moe.py

@@ -1,12 +1,11 @@
 # SPDX-License-Identifier: Apache-2.0
 """
-Test DeepEP + DeepGEMM integration 
+Test DeepEP + DeepGEMM integration
 DeepGEMM are gemm kernels specialized for the
 fp8 block-quantized case.
 """
 
 import dataclasses
-import importlib
 from typing import Optional
 
 import pytest
@@ -18,41 +17,34 @@ from vllm.config import VllmConfig, set_current_vllm_config
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
 from vllm.model_executor.layers.fused_moe.modular_kernel import (
     FusedMoEModularKernel)
-from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    per_token_group_quant_fp8)
 from vllm.platforms import current_platform
+from vllm.utils import has_deep_ep, has_deep_gemm
 
-from .utils import ProcessGroupInfo, parallel_launch
+from .parallel_utils import ProcessGroupInfo, parallel_launch
+from .utils import make_test_weights
 
-has_deep_ep = importlib.util.find_spec("deep_ep") is not None
-
-try:
-    import deep_gemm
-    has_deep_gemm = True
-except ImportError:
-    has_deep_gemm = False
-
-if has_deep_ep:
+if has_deep_ep():
     from vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize import (  # noqa: E501
         DeepEPHTPrepareAndFinalize)
     from vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize import (  # noqa: E501
         DeepEPLLPrepareAndFinalize)
 
-    from .deepep_utils import DeepEPHTArgs, DeepEPLLArgs, make_deepep_a2a
+    from .parallel_utils import DeepEPHTArgs, DeepEPLLArgs, make_deepep_a2a
+
+if has_deep_gemm():
 
-if has_deep_gemm:
     from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
         BatchedDeepGemmExperts)
     from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
         DeepGemmExperts)
 
 requires_deep_ep = pytest.mark.skipif(
-    not has_deep_ep,
+    not has_deep_ep(),
     reason="Requires deep_ep kernels",
 )
 
 requires_deep_gemm = pytest.mark.skipif(
-    not has_deep_gemm,
+    not has_deep_gemm(),
     reason="Requires deep_gemm kernels",
 )
 
@@ -66,25 +58,6 @@ def next_power_of_2(x):
     return 2**math.ceil(math.log2(x))
 
 
-def per_block_cast_to_fp8(
-        x: torch.Tensor,
-        block_size_n: int = 128) -> tuple[torch.Tensor, torch.Tensor]:
-    assert x.dim() == 2
-    m, n = x.shape
-    x_padded = torch.zeros(
-        (deep_gemm.ceil_div(m, 128) * 128,
-         deep_gemm.ceil_div(n, block_size_n) * block_size_n),
-        dtype=x.dtype,
-        device=x.device)
-    x_padded[:m, :n] = x
-    x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, block_size_n)
-    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
-    x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
-    x_scaled_sub = x_scaled.view_as(x_padded)[:m, :n].contiguous()
-    scales = (x_amax / 448.0).view(x_view.size(0), x_view.size(2))
-    return x_scaled_sub, scales
-
-
 def make_block_quant_fp8_weights(
     e: int,
     n: int,
@@ -92,43 +65,11 @@ def make_block_quant_fp8_weights(
     block_size: list[int],
 ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
     """
-    Return weights w1, w2, w1q, w2q, w1_scale, w2_scale
+    Return weights w1q, w2q, w1_scale, w2_scale
     """
-    dtype = torch.bfloat16
-
-    fp8_info = torch.finfo(torch.float8_e4m3fn)
-    fp8_max, fp8_min = fp8_info.max, fp8_info.min
-
-    w1_bf16 = torch.randn((e, 2 * n, k), dtype=dtype) / 10
-    w1_bf16 = w1_bf16.clamp(min=fp8_min, max=fp8_max).to(dtype=dtype)
-
-    w2_bf16 = torch.randn((e, k, n), dtype=dtype) / 10
-    w2_bf16 = w2_bf16.clamp(min=fp8_min, max=fp8_max).to(dtype=dtype)
-
-    block_n, block_k = block_size[0], block_size[1]
-    n_tiles_w1 = ((2 * n) + block_n - 1) // block_n
-    k_tiles_w1 = (k + block_k - 1) // block_k
-    n_tiles_w2 = (k + block_n - 1) // block_n
-    k_tiles_w2 = (n + block_k - 1) // block_k
-
-    w1 = torch.empty_like(w1_bf16, dtype=torch.float8_e4m3fn)
-    w2 = torch.empty_like(w2_bf16, dtype=torch.float8_e4m3fn)
-
-    w1_s = torch.empty((e, n_tiles_w1, k_tiles_w1),
-                       device="cuda",
-                       dtype=torch.float32)
-    w2_s = torch.empty((e, n_tiles_w2, k_tiles_w2),
-                       device="cuda",
-                       dtype=torch.float32)
-
-    assert w1_s.shape == (e, (2 * n + 127) // 128, (k + 127) // 128)
-    assert (w2.shape[-2] + block_n - 1) // block_n == w2_s.shape[-2]
-
-    for i in range(e):
-        w1[i], w1_s[i] = per_block_cast_to_fp8(w1_bf16[i])
-        w2[i], w2_s[i] = per_block_cast_to_fp8(w2_bf16[i])
-
-    return w1, w2, w1_s, w2_s
+    w1, w1q, w1_scale, w2, w2q, w2_scale = make_test_weights(
+        e, n, k, torch.bfloat16, torch.float8_e4m3fn, block_size)
+    return w1q, w2q, w1_scale, w2_scale
 
 
 @dataclasses.dataclass
@@ -138,6 +79,7 @@ class TestConfig:
     k: int
     n: int
     num_experts: int
+    per_act_token_quant: bool
     block_size: list[int]
     # configs for testing low-latency kernels
     low_latency: bool
@@ -156,8 +98,7 @@ class TestTensors:
     def make(config: TestConfig, rank) -> "TestTensors":
 
         dtype = torch.bfloat16
-        topk, m, k, block_size = (config.topk, config.m, config.k,
-                                  config.block_size)
+        topk, m, k = (config.topk, config.m, config.k)
 
         fp8_info = torch.finfo(torch.float8_e4m3fn)
         fp8_max, fp8_min = fp8_info.max, fp8_info.min
@@ -165,9 +106,7 @@ class TestTensors:
         rank_tokens = torch.randn(
             (m, k), device=torch.cuda.current_device(), dtype=dtype) / 10.0
         rank_tokens = rank_tokens.clamp(min=fp8_min, max=fp8_max)
-
-        block_k = block_size[1]
-        _, rank_token_scales = per_token_group_quant_fp8(rank_tokens, block_k)
+        rank_token_scales = None
 
         topk_ids = torch.randint(
             low=0,
@@ -207,10 +146,12 @@ def make_ll_modular_kernel(pg: ProcessGroup, pgi: ProcessGroupInfo,
         q_dtype=q_dtype,
         block_shape=test_config.block_size)
 
-    fused_experts = BatchedDeepGemmExperts(max_num_tokens=max_tokens_per_rank,
-                                           world_size=pgi.world_size,
-                                           dp_size=dp_size,
-                                           block_shape=test_config.block_size)
+    fused_experts = BatchedDeepGemmExperts(
+        max_num_tokens=max_tokens_per_rank,
+        world_size=pgi.world_size,
+        dp_size=dp_size,
+        block_shape=test_config.block_size,
+        per_act_token_quant=test_config.per_act_token_quant)
     mk = FusedMoEModularKernel(prepare_finalize=a2a,
                                fused_experts=fused_experts)
     return mk
@@ -432,6 +373,7 @@ def test_ht_deepep_deepgemm_moe(mnk: tuple[int, int, int], num_experts: int,
     """
     Tests for High-Throughput DeepEP + DeepGemm integration.
     """
+    import deep_gemm
 
     m, n, k = mnk
     current_platform.seed_everything(7)
@@ -448,6 +390,7 @@ def test_ht_deepep_deepgemm_moe(mnk: tuple[int, int, int], num_experts: int,
                         k=k,
                         n=n,
                         num_experts=num_experts,
+                        per_act_token_quant=False,
                         block_size=block_size,
                         low_latency=False,
                         use_fp8_dispatch=None)
@@ -480,10 +423,14 @@ USE_FP8_DISPATCH = [False]
 @pytest.mark.parametrize("world_dp_size", [(2, 1)])
 @requires_deep_ep
 @requires_deep_gemm
-def test_ll_deepep_deepgemm_moe(mnk: tuple[int, int,
-                                           int], num_experts: int, topk: int,
-                                use_fp8_dispatch: bool, block_size: list[int],
-                                world_dp_size: tuple[int, int]):
+def test_ll_deepep_deepgemm_moe(
+    mnk: tuple[int, int, int],
+    num_experts: int,
+    topk: int,
+    use_fp8_dispatch: bool,
+    block_size: list[int],
+    world_dp_size: tuple[int, int],
+):
     """
     Tests for Low-Latency DeepEP + DeepGemm integration.
     """
@@ -501,6 +448,7 @@ def test_ll_deepep_deepgemm_moe(mnk: tuple[int, int,
         k=k,
         n=n,
         num_experts=num_experts,
+        per_act_token_quant=False,
         block_size=block_size,
         low_latency=True,
         use_fp8_dispatch=use_fp8_dispatch,

tests/kernels/moe/test_deepep_moe.py

@@ -4,7 +4,6 @@ Test deepep dispatch-combine logic
 """
 
 import dataclasses
-import importlib
 from typing import Optional, Union
 
 import pytest
@@ -22,21 +21,20 @@ from vllm.model_executor.layers.fused_moe.modular_kernel import (
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8)
 from vllm.platforms import current_platform
+from vllm.utils import has_deep_ep
 
-from .utils import ProcessGroupInfo, parallel_launch
+from .parallel_utils import ProcessGroupInfo, parallel_launch
 
-has_deep_ep = importlib.util.find_spec("deep_ep") is not None
-
-if has_deep_ep:
+if has_deep_ep():
     from vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize import (  # noqa: E501
         DeepEPHTPrepareAndFinalize)
     from vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize import (  # noqa: E501
         DeepEPLLPrepareAndFinalize)
 
-    from .deepep_utils import DeepEPHTArgs, DeepEPLLArgs, make_deepep_a2a
+    from .parallel_utils import DeepEPHTArgs, DeepEPLLArgs, make_deepep_a2a
 
 requires_deep_ep = pytest.mark.skipif(
-    not has_deep_ep,
+    not has_deep_ep(),
     reason="Requires deep_ep kernels",
 )
 
@@ -104,10 +102,6 @@ class TestTensors:
         rank_tokens = torch.randn(
             (config.m, config.k), device="cuda", dtype=token_dtype) / 10
         rank_token_scales = None
-        if config.dtype == torch.float8_e4m3fn:
-            # low_latency_mode kernels dont support per-token quant.
-            _, rank_token_scales = ops.scaled_fp8_quant(
-                rank_tokens, use_per_token_if_dynamic=not low_latency_mode)
 
         topk = torch.randint(low=0,
                              high=config.num_experts,
@@ -123,11 +117,18 @@ class TestTensors:
                            config=config)
 
 
-def make_modular_kernel(pg: ProcessGroup, pgi: ProcessGroupInfo,
-                        low_latency_mode: bool, hidden_size: int, dp_size: int,
-                        num_experts: int, num_local_experts: int,
-                        q_dtype: Optional[torch.dtype],
-                        use_fp8_dispatch: bool) -> FusedMoEModularKernel:
+def make_modular_kernel(
+    pg: ProcessGroup,
+    pgi: ProcessGroupInfo,
+    low_latency_mode: bool,
+    hidden_size: int,
+    dp_size: int,
+    num_experts: int,
+    num_local_experts: int,
+    q_dtype: Optional[torch.dtype],
+    use_fp8_dispatch: bool,
+    per_act_token_quant: bool,
+) -> FusedMoEModularKernel:
 
     is_quantized = q_dtype is not None
 
@@ -154,6 +155,7 @@ def make_modular_kernel(pg: ProcessGroup, pgi: ProcessGroupInfo,
                         deepep_ll_args = ll_args)
 
     if low_latency_mode:
+        assert not per_act_token_quant, "not supported in ll mode"
         fused_experts = BatchedTritonExperts(
             max_num_tokens=MAX_TOKENS_PER_RANK,
             world_size=pgi.world_size,
@@ -161,25 +163,37 @@ def make_modular_kernel(pg: ProcessGroup, pgi: ProcessGroupInfo,
             use_fp8_w8a8=is_quantized,
             use_int8_w8a8=False,
             use_int8_w8a16=False,
-            use_int4_w4a16=False)
+            use_int4_w4a16=False,
+            per_act_token_quant=False,
+        )
     else:
-        fused_experts = TritonExperts(use_fp8_w8a8=is_quantized,
-                                      use_int8_w8a8=False,
-                                      use_int8_w8a16=False,
-                                      use_int4_w4a16=False,
-                                      per_channel_quant=False)
+        fused_experts = TritonExperts(
+            use_fp8_w8a8=is_quantized,
+            use_int8_w8a8=False,
+            use_int8_w8a16=False,
+            use_int4_w4a16=False,
+            per_act_token_quant=per_act_token_quant,
+        )
 
     mk = FusedMoEModularKernel(prepare_finalize=a2a,
                                fused_experts=fused_experts)
     return mk
 
 
-def deep_ep_moe_impl(pg: ProcessGroup, pgi: ProcessGroupInfo,
-                     low_latency_mode: bool, dp_size: int,
-                     test_tensors: TestTensors, w1: torch.Tensor,
-                     w2: torch.Tensor, w1_scale: Optional[torch.Tensor],
-                     w2_scale: Optional[torch.Tensor], num_experts: int,
-                     use_fp8_dispatch: bool) -> torch.Tensor:
+def deep_ep_moe_impl(
+    pg: ProcessGroup,
+    pgi: ProcessGroupInfo,
+    low_latency_mode: bool,
+    dp_size: int,
+    test_tensors: TestTensors,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    w1_scale: Optional[torch.Tensor],
+    w2_scale: Optional[torch.Tensor],
+    num_experts: int,
+    use_fp8_dispatch: bool,
+    per_act_token_quant: bool,
+) -> torch.Tensor:
 
     num_local_experts = w1.size(0)
 
@@ -201,11 +215,9 @@ def deep_ep_moe_impl(pg: ProcessGroup, pgi: ProcessGroupInfo,
         q_dtype = torch.float8_e4m3fn
 
     # Make modular kernel
-    mk: FusedMoEModularKernel = make_modular_kernel(pg, pgi, low_latency_mode,
-                                                    hidden_size, dp_size,
-                                                    num_experts,
-                                                    num_local_experts, q_dtype,
-                                                    use_fp8_dispatch)
+    mk: FusedMoEModularKernel = make_modular_kernel(
+        pg, pgi, low_latency_mode, hidden_size, dp_size, num_experts,
+        num_local_experts, q_dtype, use_fp8_dispatch, per_act_token_quant)
 
     out_hidden_states = torch.empty_like(test_tensors.rank_tokens)
     total_num_tokens = test_tensors.rank_tokens.size(0)
@@ -259,9 +271,15 @@ def deep_ep_moe_impl(pg: ProcessGroup, pgi: ProcessGroupInfo,
     return out_hidden_states
 
 
-def torch_moe_impl(test_tensors: TestTensors, w1: torch.Tensor,
-                   w2: torch.Tensor, w1_scale: Optional[torch.Tensor],
-                   w2_scale: Optional[torch.Tensor], using_fp8_dispatch: bool):
+def torch_moe_impl(
+    test_tensors: TestTensors,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    w1_scale: Optional[torch.Tensor],
+    w2_scale: Optional[torch.Tensor],
+    using_fp8_dispatch: bool,
+    per_act_token_quant: bool,
+):
 
     a, topk_ids, topk_weights = (test_tensors.rank_tokens, test_tensors.topk,
                                  test_tensors.topk_weights)
@@ -269,6 +287,7 @@ def torch_moe_impl(test_tensors: TestTensors, w1: torch.Tensor,
         # The DeepEP implementation is requested to dispatch using FP8.
         # For numerical stability for testing, emulate the fp8 dispatch by
         # blockwise quant and de-quant.
+        assert not per_act_token_quant
         a = test_tensors.rank_tokens
         aq, aq_scale = per_token_group_quant_fp8(a, 128)
         a = (aq.view(-1, 128).to(torch.float32) * aq_scale.view(-1, 1)).view(
@@ -312,6 +331,7 @@ def _deep_ep_moe(
     w1_scale: Optional[torch.Tensor],
     w2_scale: Optional[torch.Tensor],
     use_fp8_dispatch: bool,
+    per_act_token_quant: bool,
 ):
 
     if not low_latency_mode:
@@ -333,7 +353,8 @@ def _deep_ep_moe(
     with set_current_vllm_config(VllmConfig()):
         # Reference
         torch_combined = torch_moe_impl(test_tensors, w1, w2, w1_scale,
-                                        w2_scale, use_fp8_dispatch)
+                                        w2_scale, use_fp8_dispatch,
+                                        per_act_token_quant)
 
         # Splice experts for this rank.
         num_local_experts = config.num_experts // pgi.world_size
@@ -358,6 +379,7 @@ def _deep_ep_moe(
             w2_scale_ep,
             config.num_experts,
             use_fp8_dispatch,
+            per_act_token_quant,
         )
 
     torch.testing.assert_close(
@@ -386,10 +408,16 @@ DTYPES = [torch.bfloat16, torch.float8_e4m3fn]
 @pytest.mark.parametrize("num_experts", [32])
 @pytest.mark.parametrize("topk", [6])
 @pytest.mark.parametrize("world_dp_size", [(2, 1)])
+@pytest.mark.parametrize("per_act_token_quant", [False, True])
 @requires_deep_ep
-def test_deep_ep_moe(dtype: torch.dtype, mnk: tuple[int, int, int],
-                     num_experts: int, topk: int, world_dp_size: tuple[int,
-                                                                       int]):
+def test_deep_ep_moe(
+    dtype: torch.dtype,
+    mnk: tuple[int, int, int],
+    num_experts: int,
+    topk: int,
+    world_dp_size: tuple[int, int],
+    per_act_token_quant: bool,
+):
     low_latency_mode = False
     use_fp8_dispatch = False
     m, n, k = mnk
@@ -406,7 +434,8 @@ def test_deep_ep_moe(dtype: torch.dtype, mnk: tuple[int, int, int],
     w1, w2, w1_scale, w2_scale = make_weights(num_experts, n, k, dtype)
 
     parallel_launch(world_size, _deep_ep_moe, low_latency_mode, dp_size,
-                    config, w1, w2, w1_scale, w2_scale, use_fp8_dispatch)
+                    config, w1, w2, w1_scale, w2_scale, use_fp8_dispatch,
+                    per_act_token_quant)
 
 
 MNKs = [
@@ -456,4 +485,5 @@ def test_low_latency_deep_ep_moe(dtype: torch.dtype, mnk: tuple[int, int, int],
     w1, w2, w1_scale, w2_scale = make_weights(num_experts, n, k, dtype)
 
     parallel_launch(world_size, _deep_ep_moe, low_latency_mode, dp_size,
-                    config, w1, w2, w1_scale, w2_scale, use_fp8_dispatch)
+                    config, w1, w2, w1_scale, w2_scale, use_fp8_dispatch,
+                    False)

tests/kernels/moe/test_deepgemm.py

@@ -0,0 +1,225 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Unit-test DeepGEMM FP8 kernels (no DeepEP).
+Compare DeepGEMM path against the Triton fallback inside vLLM's fused_experts.
+"""
+
+import importlib
+import math
+
+import pytest
+import torch
+
+# vLLM fused-expert reference (Triton fallback + DeepGEMM option)
+from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    per_token_group_quant_fp8)
+from vllm.utils import cdiv
+
+has_deep_gemm = importlib.util.find_spec("deep_gemm") is not None
+
+if has_deep_gemm:
+    import deep_gemm
+    BLOCK_M = deep_gemm.get_m_alignment_for_contiguous_layout()
+    BLOCK_SIZE = [BLOCK_M, BLOCK_M]
+
+requires_deep_gemm = pytest.mark.skipif(
+    not has_deep_gemm,
+    reason="Requires deep_gemm kernels",
+)
+
+
+def calc_diff(x: torch.Tensor, y: torch.Tensor):
+    x, y = x.double(), y.double()
+    denominator = (x * x + y * y).sum()
+    sim = 2 * (x * y).sum() / denominator
+    return 1 - sim
+
+
+def per_block_cast_to_fp8(
+        x: torch.Tensor,
+        block_size_n: int = 128) -> tuple[torch.Tensor, torch.Tensor]:
+    assert x.dim() == 2
+    m, n = x.shape
+    x_padded = torch.zeros(
+        (cdiv(m, 128) * 128, cdiv(n, block_size_n) * block_size_n),
+        dtype=x.dtype,
+        device=x.device)
+    x_padded[:m, :n] = x
+    x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, block_size_n)
+    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
+    x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
+    x_scaled_sub = x_scaled.view_as(x_padded)[:m, :n].contiguous()
+    scales = (x_amax / 448.0).view(x_view.size(0), x_view.size(2))
+    return x_scaled_sub, scales
+
+
+def make_block_quant_fp8_weights(
+    e: int,
+    n: int,
+    k: int,
+    block_size: list[int],
+):
+    """
+    Generate (w1, w2) expert weights and their per-block scale tensors
+    in FP8 block-quantized format.
+
+      w1 shape: (E, 2N, K)
+      w2 shape: (E, K, N)
+    """
+    dtype = torch.bfloat16
+    fp8_max, fp8_min = torch.finfo(torch.float8_e4m3fn).max, torch.finfo(
+        torch.float8_e4m3fn).min
+
+    # bf16 reference weights
+    w1_bf16 = torch.randn(e, 2 * n, k, device="cuda", dtype=dtype) / 10
+    w2_bf16 = torch.randn(e, k, n, device="cuda", dtype=dtype) / 10
+    w1_bf16.clamp_(fp8_min, fp8_max)
+    w2_bf16.clamp_(fp8_min, fp8_max)
+
+    block_n, block_k = block_size
+    n_tiles_w1 = math.ceil((2 * n) / block_n)
+    k_tiles_w1 = math.ceil(k / block_k)
+    n_tiles_w2 = math.ceil(k / block_n)
+    k_tiles_w2 = math.ceil(n / block_k)
+
+    w1 = torch.empty_like(w1_bf16, dtype=torch.float8_e4m3fn)
+    w2 = torch.empty_like(w2_bf16, dtype=torch.float8_e4m3fn)
+    w1_s = torch.empty(e,
+                       n_tiles_w1,
+                       k_tiles_w1,
+                       device="cuda",
+                       dtype=torch.float32)
+    w2_s = torch.empty(e,
+                       n_tiles_w2,
+                       k_tiles_w2,
+                       device="cuda",
+                       dtype=torch.float32)
+
+    for i in range(e):
+        w1[i], w1_s[i] = per_block_cast_to_fp8(w1_bf16[i])
+        w2[i], w2_s[i] = per_block_cast_to_fp8(w2_bf16[i])
+
+    return w1, w2, w1_s, w2_s
+
+
+def run_single_case(m, n, k, topk, num_experts, block_size):
+    """
+    Run one (M,N,K) configuration on a single GPU and assert DeepGEMM ==
+    Triton baseline within tolerance.
+    """
+    tokens_bf16 = torch.randn(
+        m, k, device="cuda", dtype=torch.bfloat16).clamp_min_(-1).clamp_max_(1)
+    _, a1_scale = per_token_group_quant_fp8(tokens_bf16, block_size[1])
+
+    # expert weight tensors
+    w1, w2, w1_s, w2_s = make_block_quant_fp8_weights(num_experts, n, k,
+                                                      block_size)
+
+    router_logits = torch.randn(m,
+                                num_experts,
+                                device="cuda",
+                                dtype=torch.float32)
+    topk_weights, topk_ids = torch.topk(router_logits, k=topk, dim=-1)
+    topk_weights = torch.nn.functional.softmax(topk_weights, dim=-1)
+
+    # triton referrence
+    out_triton = fused_experts(
+        hidden_states=tokens_bf16,
+        w1=w1,
+        w2=w2,
+        topk_weights=topk_weights,
+        topk_ids=topk_ids,
+        inplace=False,
+        use_fp8_w8a8=True,
+        w1_scale=w1_s,
+        w2_scale=w2_s,
+        a1_scale=a1_scale,
+        block_shape=block_size,
+        allow_deep_gemm=False,
+    )
+
+    # DeepGemm
+    out_deepgemm = fused_experts(
+        hidden_states=tokens_bf16,
+        w1=w1,
+        w2=w2,
+        topk_weights=topk_weights,
+        topk_ids=topk_ids,
+        inplace=False,
+        use_fp8_w8a8=True,
+        w1_scale=w1_s,
+        w2_scale=w2_s,
+        a1_scale=a1_scale,
+        block_shape=block_size,
+        allow_deep_gemm=True,
+    )
+
+    base = out_triton.abs().mean()
+    atol = 0.1 * base.clamp(min=1e-2)  # 10% of mean, but not lower than 1e-3
+    rtol = 0.05
+    # ----- Compare -----
+    torch.testing.assert_close(
+        out_deepgemm.to(torch.float32),
+        out_triton.to(torch.float32),
+        rtol=rtol,
+        atol=float(atol),
+    )
+
+
+# Note: W1 has shape (E, 2N, K), so N = 512
+# can trigger the deepgemm path.
+MNKs = [
+    (1024, 512, 128),
+    (1024, 512, 512),
+    (2048, 512, 512),
+    (512, 1024, 1024),
+    (512, 2048, 2048),
+    (4096, 4096, 1024),
+]
+
+TOPKS = [2, 6]
+NUM_EXPERTS = [32]
+
+
+@pytest.mark.parametrize("mnk", MNKs)
+@pytest.mark.parametrize("topk", TOPKS)
+@pytest.mark.parametrize("num_experts", NUM_EXPERTS)
+@requires_deep_gemm
+def test_deepgemm_vs_triton(mnk, topk, num_experts, monkeypatch):
+
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_USE_DEEP_GEMM", "1")
+
+        _fused_moe_mod = importlib.import_module(
+            "vllm.model_executor.layers.fused_moe.fused_moe")
+
+        call_counter = {"cnt": 0}
+
+        orig_fn = _fused_moe_mod.deep_gemm_moe_fp8
+
+        def _spy_deep_gemm_moe_fp8(*args, **kwargs):
+            call_counter["cnt"] += 1
+            return orig_fn(*args, **kwargs)
+
+        monkeypatch.setattr(_fused_moe_mod, "deep_gemm_moe_fp8",
+                            _spy_deep_gemm_moe_fp8)
+
+        m, n, k = mnk
+
+        if topk > num_experts:
+            pytest.skip(f"topk={topk} > num_experts={num_experts}")
+
+        run_single_case(
+            m=m,
+            n=n,
+            k=k,
+            topk=topk,
+            num_experts=num_experts,
+            block_size=BLOCK_SIZE,
+        )
+
+        # ensure that the DeepGEMM path was indeed taken.
+        assert call_counter["cnt"] == 1, \
+            f"DeepGEMM path was not executed during the test. " \
+            f"Call counter: {call_counter['cnt']}"

tests/kernels/moe/test_moe.py

@@ -17,6 +17,7 @@ from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
 import vllm.model_executor.layers.fused_moe  # noqa
 from tests.kernels.utils import opcheck, stack_and_dev, torch_moe
 from vllm.config import VllmConfig, set_current_vllm_config
+from vllm.distributed.parallel_state import init_distributed_environment
 from vllm.forward_context import set_forward_context
 from vllm.model_executor.layers.fused_moe import fused_moe
 from vllm.model_executor.layers.fused_moe.fused_moe import (
@@ -142,6 +143,10 @@ def test_fused_moe(
     # Setup test data
     #
 
+    #
+    # Setup test data
+    #
+
     a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
     w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
     w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10
@@ -169,7 +174,7 @@ def test_fused_moe(
                                               use_int8_w8a8=False,
                                               use_int8_w8a16=False,
                                               use_int4_w4a16=False,
-                                              per_channel_quant=False,
+                                              per_act_token_quant=False,
                                               block_shape=None)
 
     def m_fused_moe(
@@ -365,6 +370,13 @@ def test_mixtral_moe(dtype: torch.dtype, padding: bool, use_rocm_aiter: bool,
         if dtype == torch.float32:
             pytest.skip("AITER ROCm test skip for float32")
 
+    monkeypatch.setenv('RANK', "0")
+    monkeypatch.setenv('LOCAL_RANK', "0")
+    monkeypatch.setenv('WORLD_SIZE', "1")
+    monkeypatch.setenv('MASTER_ADDR', 'localhost')
+    monkeypatch.setenv('MASTER_PORT', '12345')
+    init_distributed_environment()
+
     # Instantiate our and huggingface's MoE blocks
     vllm_config.compilation_config.static_forward_context = dict()
     with (set_current_vllm_config(vllm_config),

tests/kernels/moe/test_nvfp4_moe.py

@@ -14,7 +14,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
 from vllm.platforms import current_platform
 
 if not current_platform.has_device_capability(100):
-    pytest.skip(reason="Nvfp4 Requires compute capability of 10 or above.",
+    pytest.skip("Nvfp4 Requires compute capability of 10 or above.",
                 allow_module_level=True)
 
 MNK_FACTORS = [

tests/kernels/moe/test_pplx_cutlass_moe.py

@@ -15,7 +15,7 @@ from vllm.model_executor.layers.fused_moe.modular_kernel import (
     FusedMoEModularKernel)
 from vllm.platforms import current_platform
 
-from .utils import ProcessGroupInfo, parallel_launch
+from .parallel_utils import ProcessGroupInfo, parallel_launch
 
 try:
     from pplx_kernels import AllToAll
@@ -93,7 +93,7 @@ def pplx_cutlass_moe(
         num_experts=num_experts,
         experts_per_token=topk,
         rank=rank,
-        world_size=pgi.world_size,
+        world_size=world_size,
         dp_size=dp_size,
         hidden_dim=hidden_dim,
         hidden_dim_bytes=hidden_dim,  # because a.dtype.itemsize == 1
@@ -118,8 +118,6 @@ def pplx_cutlass_moe(
         pgi.world_size,
         rank,
         dp_size,
-        quant_dtype=torch.float8_e4m3fn,
-        per_act_token=per_act_token,
     )
 
     experts = CutlassExpertsFp8((num_experts + world_size - 1) // world_size,

tests/kernels/moe/test_pplx_moe.py

@@ -18,18 +18,20 @@ try:
 except ImportError:
     has_pplx = False
 
+from tests.kernels.moe.utils import make_test_weights, naive_batched_moe
 from tests.kernels.utils import torch_experts
 from vllm.config import VllmConfig, set_current_vllm_config
-from vllm.model_executor.layers.fused_moe import override_config
+from vllm.model_executor.layers.fused_moe import fused_topk, override_config
+from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
 from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
-    BatchedExperts, BatchedPrepareAndFinalize, BatchedTritonExperts)
-from vllm.model_executor.layers.fused_moe.fused_moe import (fused_topk,
-                                                            get_default_config)
+    BatchedPrepareAndFinalize, BatchedTritonExperts, NaiveBatchedExperts)
+from vllm.model_executor.layers.fused_moe.fused_moe import get_default_config
 from vllm.model_executor.layers.fused_moe.modular_kernel import (
     FusedMoEModularKernel)
 from vllm.platforms import current_platform
+from vllm.utils import round_up
 
-from .utils import ProcessGroupInfo, parallel_launch
+from .parallel_utils import ProcessGroupInfo, parallel_launch
 
 requires_pplx = pytest.mark.skipif(
     not has_pplx,
@@ -144,25 +146,6 @@ def torch_batched_moe(
     return torch_finalize(out, topk_weight, topk_ids)
 
 
-def batched_moe(
-    a: torch.Tensor,
-    w1: torch.Tensor,
-    w2: torch.Tensor,
-    topk_weight: torch.Tensor,
-    topk_ids: torch.Tensor,
-) -> torch.Tensor:
-    num_experts = w1.shape[0]
-
-    fused_experts = FusedMoEModularKernel(
-        BatchedPrepareAndFinalize(max_num_tokens=a.shape[0],
-                                  world_size=1,
-                                  dp_size=1,
-                                  rank=0),
-        BatchedExperts(max_num_tokens=a.shape[0], dp_size=1, world_size=1))
-
-    return fused_experts(a, w1, w2, topk_weight, topk_ids, num_experts)
-
-
 @pytest.mark.parametrize("m", [1, 33, 64, 222])
 @pytest.mark.parametrize("n", [128, 1024, 2048])
 @pytest.mark.parametrize("k", [128, 512, 1024])
@@ -188,7 +171,7 @@ def test_fused_moe_batched_experts(
         topk_weight, topk_ids, _ = fused_topk(a, score, topk, False)
         baseline_output = torch_experts(a, w1, w2, topk_weight, topk_ids)
         torch_output = torch_batched_moe(a, w1, w2, topk_weight, topk_ids)
-        batched_output = batched_moe(a, w1, w2, topk_weight, topk_ids)
+        batched_output = naive_batched_moe(a, w1, w2, topk_weight, topk_ids)
 
     torch.testing.assert_close(baseline_output,
                                torch_output,
@@ -226,7 +209,6 @@ def pplx_prepare_finalize(
 
     topk = topk_ids.shape[1]
     num_tokens, hidden_dim = a.shape
-    block_size = 128
     device = pgi.device
     rank = pgi.rank
     world_size = pgi.world_size
@@ -241,9 +223,7 @@ def pplx_prepare_finalize(
         dp_size=dp_size,
         hidden_dim=hidden_dim,
         hidden_dim_bytes=hidden_dim * a.dtype.itemsize,
-        hidden_dim_scale_bytes=(0 if a.dtype.itemsize != 1 else
-                                ((hidden_dim + block_size - 1) // block_size *
-                                 torch.float32.itemsize)),
+        hidden_dim_scale_bytes=0,
     )
 
     if group_name is None:
@@ -260,7 +240,6 @@ def pplx_prepare_finalize(
         world_size,
         rank,
         dp_size,
-        a.dtype,
     )
 
     a_chunk = chunk_by_rank(a, rank, world_size).to(device)
@@ -276,6 +255,7 @@ def pplx_prepare_finalize(
         num_experts,
         None,
         False,
+        FusedMoEQuantConfig(),
     )
 
     b_a = b_a * 1.5
@@ -350,6 +330,7 @@ def _pplx_prepare_finalize(
 # TODO (bnell): this test point does not work for odd M due to how the test is
 # written, not due to limitations of the pplx kernels.  The pplx_moe
 # test below is able to deal with odd M.
+# TODO (bnell) add fp8 tests
 @pytest.mark.parametrize("mnk", PPLX_PREPARE_COMBOS)
 @pytest.mark.parametrize("e", NUM_EXPERTS)
 @pytest.mark.parametrize("topk", TOP_KS)
@@ -386,18 +367,31 @@ def pplx_moe(
     w2: torch.Tensor,
     topk_weight: torch.Tensor,
     topk_ids: torch.Tensor,
+    w1_scale: Optional[torch.Tensor] = None,
+    w2_scale: Optional[torch.Tensor] = None,
+    qtype: Optional[torch.dtype] = None,
+    per_act_token_quant=False,
+    block_shape: Optional[list[int]] = None,
     use_compile: bool = False,
     use_cudagraphs: bool = True,
 ) -> torch.Tensor:
     from vllm.model_executor.layers.fused_moe.pplx_prepare_finalize import (
-        PplxPrepareAndFinalize)
+        PplxPrepareAndFinalize, pplx_hidden_dim_scale_bytes)
 
     device = torch.device("cuda", rank)
     hidden_dim = a.shape[1]
     num_experts = w1.shape[0]
-    block_size = 128
     topk = topk_ids.shape[1]
-    max_num_tokens = rank_chunk(a.shape[0], 0, world_size)
+    max_num_tokens = round_up(rank_chunk(a.shape[0], 0, world_size), 64)
+
+    hidden_dim_bytes, scale_bytes = pplx_hidden_dim_scale_bytes(
+        max_num_tokens,
+        hidden_dim,
+        a.dtype,
+        qtype,
+        per_act_token_quant=per_act_token_quant,
+        block_shape=block_shape,
+    )
 
     args = dict(
         max_num_tokens=max_num_tokens,
@@ -407,10 +401,8 @@ def pplx_moe(
         world_size=world_size,
         dp_size=dp_size,
         hidden_dim=hidden_dim,
-        hidden_dim_bytes=hidden_dim * a.dtype.itemsize,
-        hidden_dim_scale_bytes=(0 if a.dtype.itemsize != 1 else
-                                ((hidden_dim + block_size - 1) // block_size *
-                                 torch.float32.itemsize)),
+        hidden_dim_bytes=hidden_dim_bytes,
+        hidden_dim_scale_bytes=scale_bytes,
     )
 
     if group_name is None:
@@ -429,9 +421,11 @@ def pplx_moe(
         dp_size,
     )
 
-    experts = BatchedTritonExperts(max_num_tokens=a.shape[0],
+    experts = BatchedTritonExperts(max_num_tokens=max_num_tokens,
                                    world_size=world_size,
-                                   dp_size=dp_size)
+                                   dp_size=dp_size,
+                                   use_fp8_w8a8=qtype == torch.float8_e4m3fn,
+                                   block_shape=block_shape)
 
     fused_experts = FusedMoEModularKernel(
         prepare_finalize,
@@ -447,6 +441,13 @@ def pplx_moe(
     w1_chunk = chunk_by_rank(w1, rank, world_size).to(device)
     w2_chunk = chunk_by_rank(w2, rank, world_size).to(device)
 
+    if w1_scale is not None:
+        w1_scale_chunk = chunk_by_rank(w1_scale, rank, world_size).to(device)
+        w2_scale_chunk = chunk_by_rank(w2_scale, rank, world_size).to(device)
+    else:
+        w1_scale_chunk = None
+        w2_scale_chunk = None
+
     # Note: for now use_compile will error out if the problem size is
     # large enough to trigger chunking. I'm leaving the flag and
     # setup code in case we are able to revisit this later.
@@ -465,6 +466,8 @@ def pplx_moe(
                          w2_chunk,
                          chunk_topk_weight,
                          chunk_topk_ids,
+                         w1_scale=w1_scale_chunk,
+                         w2_scale=w2_scale_chunk,
                          global_num_experts=num_experts)
 
     if use_cudagraphs:
@@ -477,6 +480,8 @@ def pplx_moe(
                                  w2_chunk,
                                  chunk_topk_weight,
                                  chunk_topk_ids,
+                                 w1_scale=w1_scale_chunk,
+                                 w2_scale=w2_scale_chunk,
                                  global_num_experts=num_experts)
 
         torch.cuda.synchronize()
@@ -505,9 +510,9 @@ def _batched_moe(pgi, dp_size, a, w1, w2, topk_weight, topk_ids):
         rank=rank,
     )
 
-    experts = BatchedExperts(max_num_tokens=a.shape[0],
-                             world_size=1,
-                             dp_size=1)
+    experts = NaiveBatchedExperts(max_num_tokens=a.shape[0],
+                                  world_size=1,
+                                  dp_size=1)
 
     fused_experts = FusedMoEModularKernel(
         prepare_finalize,
@@ -539,7 +544,12 @@ def _pplx_moe(
     w2: torch.Tensor,
     score: torch.Tensor,
     topk: int,
-    use_internode: bool,
+    w1_s: Optional[torch.Tensor] = None,
+    w2_s: Optional[torch.Tensor] = None,
+    qtype: Optional[torch.dtype] = None,
+    per_act_token_quant: bool = False,
+    block_shape: Optional[list[int]] = None,
+    use_internode: bool = False,
 ):
     if use_internode:
         uid = nvshmem_get_unique_id(
@@ -557,11 +567,28 @@ def _pplx_moe(
 
     moe_config = get_default_config(m, e, n, k, topk, a.dtype, False)
 
+    device = torch.device("cuda", pgi.rank)
+    a = a.to(device)
+    w1 = w1.to(device)
+    w2 = w2.to(device)
+    w1_s = w1_s.to(device) if w1_s is not None else None
+    w2_s = w2_s.to(device) if w2_s is not None else None
+
     with set_current_vllm_config(vllm_config), override_config(moe_config):
         topk_weight, topk_ids, _ = fused_topk(a, score, topk, False)
-        torch_output = torch_experts(a, w1, w2, topk_weight, topk_ids)
+        torch_output = torch_experts(a,
+                                     w1,
+                                     w2,
+                                     topk_weight,
+                                     topk_ids,
+                                     w1_scale=w1_s,
+                                     w2_scale=w2_s,
+                                     quant_dtype=qtype,
+                                     per_act_token_quant=per_act_token_quant,
+                                     block_shape=block_shape)
         pplx_output = pplx_moe(group_name, pgi.rank, pgi.world_size, dp_size,
-                               a, w1, w2, topk_weight, topk_ids)
+                               a, w1, w2, topk_weight, topk_ids, w1_s, w2_s,
+                               qtype, per_act_token_quant, block_shape)
         # TODO (bnell): fix + re-enable
         #batched_output = _batched_moe(pgi, dp_size, a, w1, w2, topk_weight,
         #                              topk_ids)
@@ -581,6 +608,8 @@ def _pplx_moe(
 @pytest.mark.parametrize("topk", TOP_KS)
 @pytest.mark.parametrize("dtype", [torch.bfloat16])
 @pytest.mark.parametrize("world_dp_size", [[2, 1]])
+@pytest.mark.parametrize("per_act_token_quant", [False, True])
+@pytest.mark.parametrize("block_shape", [None, [128, 128]])
 @pytest.mark.parametrize("use_internode", [False])
 @requires_pplx
 def test_pplx_moe(
@@ -589,15 +618,33 @@ def test_pplx_moe(
     topk: int,
     dtype: torch.dtype,
     world_dp_size: tuple[int, int],
+    per_act_token_quant: bool,
+    block_shape: Optional[list[int]],
     use_internode: bool,
 ):
     current_platform.seed_everything(7)
     m, n, k = mnk
     world_size, dp_size = world_dp_size
-    a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
-    w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
-    w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10
-    score = torch.randn((m, e), device="cuda", dtype=dtype)
+
+    if dtype == torch.float8_e4m3fn:
+        use_fp8_w8a8 = True
+        quant_dtype = dtype
+    else:
+        use_fp8_w8a8 = False
+        quant_dtype = None
+
+    if not use_fp8_w8a8 and per_act_token_quant and block_shape is not None:
+        pytest.skip("Skip quantization test for non-quantized type")
+
+    a = torch.randn((m, k), device="cuda", dtype=torch.bfloat16) / 10
+    score = torch.randn((m, e), device="cuda", dtype=torch.bfloat16)
+
+    _, w1, w1_s, _, w2, w2_s = make_test_weights(e,
+                                                 n,
+                                                 k,
+                                                 quant_dtype=quant_dtype,
+                                                 block_shape=block_shape)
 
     parallel_launch(world_size, _pplx_moe, dp_size, a, w1, w2, score, topk,
+                    w1_s, w2_s, quant_dtype, per_act_token_quant, block_shape,
                     use_internode)

tests/kernels/moe/utils.py

@@ -1,194 +1,249 @@
 # SPDX-License-Identifier: Apache-2.0
-"""
-DeepEP test utilities
-"""
-import dataclasses
-import importlib
-import os
-import traceback
-from typing import Callable, Optional
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Optional
 
 import torch
-from torch.distributed import ProcessGroup
-from torch.multiprocessing import (
-    spawn)  # pyright: ignore[reportPrivateImportUsage]
-from typing_extensions import Concatenate, ParamSpec
 
-from vllm.model_executor.layers.fused_moe.utils import find_free_port
-
-has_deep_ep = importlib.util.find_spec("deep_ep") is not None
-if has_deep_ep:
-    from vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize import (  # noqa: E501
-        DeepEPHTPrepareAndFinalize)
-    from vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize import (  # noqa: E501
-        DeepEPLLPrepareAndFinalize)
-
-## Parallel Processes Utils
-
-P = ParamSpec("P")
+import vllm._custom_ops as ops
+from tests.kernels.quant_utils import (per_block_cast_to_fp8,
+                                       per_block_cast_to_int8)
+from vllm.model_executor.layers.fused_moe import fused_experts
+from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
+    BatchedPrepareAndFinalize, BatchedTritonExperts, NaiveBatchedExperts)
+from vllm.model_executor.layers.fused_moe.modular_kernel import (
+    FusedMoEModularKernel)
+from vllm.model_executor.layers.fused_moe.utils import (
+    moe_kernel_quantize_input)
+from vllm.utils import round_up
 
 
-@dataclasses.dataclass
-class ProcessGroupInfo:
-    world_size: int
-    world_local_size: int
-    rank: int
-    node_rank: int
-    local_rank: int
-    device: torch.device
+def triton_moe(
+    a: torch.Tensor,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    topk_weight: torch.Tensor,
+    topk_ids: torch.Tensor,
+    w1_scale: Optional[torch.Tensor] = None,
+    w2_scale: Optional[torch.Tensor] = None,
+    a1_scale: Optional[torch.Tensor] = None,
+    a2_scale: Optional[torch.Tensor] = None,
+    quant_dtype: Optional[torch.dtype] = None,
+    per_act_token_quant=False,
+    block_shape: Optional[list[int]] = None,
+) -> torch.Tensor:
+    return fused_experts(a,
+                         w1,
+                         w2,
+                         topk_weight,
+                         topk_ids,
+                         w1_scale=w1_scale,
+                         w2_scale=w2_scale,
+                         a1_scale=a1_scale,
+                         a2_scale=a2_scale,
+                         per_channel_quant=per_act_token_quant,
+                         use_fp8_w8a8=quant_dtype == torch.float8_e4m3fn,
+                         block_shape=block_shape)
 
 
-def _worker_parallel_launch(
-    local_rank: int,
-    world_size: int,
-    world_local_size: int,
-    node_rank: int,
-    init_method: str,
-    worker: Callable[Concatenate[ProcessGroupInfo, P], None],
-    *args: P.args,
-    **kwargs: P.kwargs,
-) -> None:
-    rank = node_rank * world_local_size + local_rank
-    torch.cuda.set_device(local_rank)
-    device = torch.device("cuda", local_rank)
-    torch.distributed.init_process_group(
-        backend="cpu:gloo,cuda:nccl",
-        init_method=init_method,
-        rank=rank,
-        world_size=world_size,
-        device_id=device,
-    )
-    barrier = torch.tensor([rank], device=device)
-    torch.distributed.all_reduce(barrier)
+def batched_moe(
+    a: torch.Tensor,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    topk_weight: torch.Tensor,
+    topk_ids: torch.Tensor,
+    w1_scale: Optional[torch.Tensor] = None,
+    w2_scale: Optional[torch.Tensor] = None,
+    a1_scale: Optional[torch.Tensor] = None,
+    a2_scale: Optional[torch.Tensor] = None,
+    quant_dtype: Optional[torch.dtype] = None,
+    per_act_token_quant: bool = False,
+    block_shape: Optional[list[int]] = None,
+) -> torch.Tensor:
+    max_num_tokens = round_up(a.shape[0], 64)
 
-    try:
-        worker(
-            ProcessGroupInfo(
-                world_size=world_size,
-                world_local_size=world_local_size,
-                rank=rank,
-                node_rank=node_rank,
-                local_rank=local_rank,
-                device=device,
-            ),
-            *args,
-            **kwargs,
-        )
-    except Exception as ex:
-        print(ex)
-        traceback.print_exc()
-        raise
-    finally:
-        torch.distributed.destroy_process_group()
-
-
-def parallel_launch(
-    world_size: int,
-    worker: Callable[Concatenate[ProcessGroupInfo, P], None],
-    *args: P.args,
-    **kwargs: P.kwargs,
-) -> None:
-    assert not kwargs
-    spawn(
-        _worker_parallel_launch,
-        args=(
-            world_size,
-            world_size,
-            0,
-            f"tcp://{os.getenv('LOCALHOST', 'localhost')}:{find_free_port()}",
-            worker,
-        ) + args,
-        nprocs=world_size,
-        join=True,
+    fused_experts = FusedMoEModularKernel(
+        BatchedPrepareAndFinalize(max_num_tokens,
+                                  world_size=1,
+                                  dp_size=1,
+                                  rank=0),
+        BatchedTritonExperts(
+            max_num_tokens=max_num_tokens,
+            world_size=1,
+            dp_size=1,
+            use_fp8_w8a8=quant_dtype == torch.float8_e4m3fn,
+            per_act_token_quant=per_act_token_quant,
+            block_shape=block_shape,
+        ),
     )
 
-
-## DeepEP specific utils
+    return fused_experts(a,
+                         w1,
+                         w2,
+                         topk_weight,
+                         topk_ids,
+                         w1_scale=w1_scale,
+                         w2_scale=w2_scale,
+                         a1_scale=a1_scale,
+                         a2_scale=a2_scale)
 
 
-@dataclasses.dataclass
-class DeepEPHTArgs:
-    num_local_experts: int
+def naive_batched_moe(
+    a: torch.Tensor,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    topk_weight: torch.Tensor,
+    topk_ids: torch.Tensor,
+    w1_scale: Optional[torch.Tensor] = None,
+    w2_scale: Optional[torch.Tensor] = None,
+    a1_scale: Optional[torch.Tensor] = None,
+    a2_scale: Optional[torch.Tensor] = None,
+    quant_dtype: Optional[torch.dtype] = None,
+    per_act_token_quant: bool = False,
+    block_shape: Optional[list[int]] = None,
+) -> torch.Tensor:
+    max_num_tokens = round_up(a.shape[0], 64)
 
-
-@dataclasses.dataclass
-class DeepEPLLArgs:
-    max_tokens_per_rank: int
-    hidden_size: int
-    num_experts: int
-    use_fp8_dispatch: bool
-
-
-def make_deepep_ht_a2a(pg: ProcessGroup,
-                       pgi: ProcessGroupInfo,
-                       dp_size: int,
-                       ht_args: DeepEPHTArgs,
-                       q_dtype: Optional[torch.dtype] = None,
-                       block_shape: Optional[list[int]] = None):
-
-    import deep_ep
-
-    # high throughput a2a
-    num_nvl_bytes = 1024 * 1024 * 1024  # 1GB
-    num_rdma_bytes, low_latency_mode, num_qps_per_rank = 0, False, 1
-    buffer = deep_ep.Buffer(group=pg,
-                            num_nvl_bytes=num_nvl_bytes,
-                            num_rdma_bytes=num_rdma_bytes,
-                            low_latency_mode=low_latency_mode,
-                            num_qps_per_rank=num_qps_per_rank)
-    return DeepEPHTPrepareAndFinalize(buffer=buffer,
-                                      world_size=pgi.world_size,
-                                      rank=pgi.rank,
-                                      dp_size=dp_size,
-                                      rank_expert_offset=pgi.rank *
-                                      ht_args.num_local_experts,
-                                      quant_dtype=q_dtype,
-                                      block_shape=block_shape)
-
-
-def make_deepep_ll_a2a(pg: ProcessGroup,
-                       pgi: ProcessGroupInfo,
-                       dp_size: int,
-                       deepep_ll_args: DeepEPLLArgs,
-                       q_dtype: Optional[torch.dtype] = None,
-                       block_shape: Optional[list[int]] = None):
-
-    import deep_ep
-
-    # low-latency a2a
-    num_rdma_bytes = deep_ep.Buffer.get_low_latency_rdma_size_hint(
-        deepep_ll_args.max_tokens_per_rank, deepep_ll_args.hidden_size,
-        pgi.world_size, deepep_ll_args.num_experts)
-
-    buffer = deep_ep.Buffer(group=pg,
-                            num_rdma_bytes=num_rdma_bytes,
-                            low_latency_mode=True,
-                            num_qps_per_rank=deepep_ll_args.num_experts //
-                            pgi.world_size)
-
-    return DeepEPLLPrepareAndFinalize(
-        buffer=buffer,
-        world_size=pgi.world_size,
-        dp_size=dp_size,
-        max_tokens_per_rank=deepep_ll_args.max_tokens_per_rank,
-        quant_dtype=q_dtype,
-        block_shape=block_shape,
-        use_fp8_dispatch=deepep_ll_args.use_fp8_dispatch,
+    fused_experts = FusedMoEModularKernel(
+        BatchedPrepareAndFinalize(max_num_tokens,
+                                  world_size=1,
+                                  dp_size=1,
+                                  rank=0),
+        NaiveBatchedExperts(
+            max_num_tokens=max_num_tokens,
+            dp_size=1,
+            world_size=1,
+            use_fp8_w8a8=quant_dtype == torch.float8_e4m3fn,
+            per_act_token_quant=per_act_token_quant,
+            block_shape=block_shape,
+        ),
     )
 
+    return fused_experts(a,
+                         w1,
+                         w2,
+                         topk_weight,
+                         topk_ids,
+                         w1_scale=w1_scale,
+                         w2_scale=w2_scale,
+                         a1_scale=a1_scale,
+                         a2_scale=a2_scale)
 
-def make_deepep_a2a(pg: ProcessGroup,
-                    pgi: ProcessGroupInfo,
-                    dp_size: int,
-                    deepep_ht_args: Optional[DeepEPHTArgs],
-                    deepep_ll_args: Optional[DeepEPLLArgs],
-                    q_dtype: Optional[torch.dtype] = None,
-                    block_shape: Optional[list[int]] = None):
-    if deepep_ht_args is not None:
-        assert deepep_ll_args is None
-        return make_deepep_ht_a2a(pg, pgi, dp_size, deepep_ht_args, q_dtype,
-                                  block_shape)
 
-    assert deepep_ll_args is not None
-    return make_deepep_ll_a2a(pg, pgi, dp_size, deepep_ll_args, q_dtype,
-                              block_shape)
+def chunk_scales(scales: Optional[torch.Tensor], start: int,
+                 end: int) -> Optional[torch.Tensor]:
+    if scales is not None:
+        if scales.numel() == 1:
+            return scales
+        else:
+            return scales[start:end]
+    return None
+
+
+def make_quantized_test_activations(
+    E: int,
+    m: int,
+    k: int,
+    in_dtype: torch.dtype,
+    quant_dtype: Optional[torch.dtype] = None,
+    block_shape: Optional[list[int]] = None,
+    per_act_token_quant: bool = False,
+) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+    a = torch.randn((E, m, k), device="cuda", dtype=in_dtype) / 10
+    a_q = a
+    a_scale = None
+
+    if quant_dtype is not None:
+        assert (quant_dtype == torch.float8_e4m3fn
+                or quant_dtype == torch.int8), "only fp8/int8 supported"
+        a_q = torch.zeros_like(a, dtype=quant_dtype)
+        a_scale_l = [None] * E
+        for e in range(E):
+            a_q[e], a_scale_l[e] = moe_kernel_quantize_input(
+                a[e], None, quant_dtype, per_act_token_quant, block_shape)
+        a_scale = torch.stack(a_scale_l)
+
+        if not per_act_token_quant and block_shape is None:
+            a_scale = a_scale.view(E, 1, 1)
+
+    return a, a_q, a_scale
+
+
+def moe_quantize_weights(
+    w: torch.Tensor,
+    w_s: Optional[torch.Tensor],
+    quant_dtype: Optional[torch.dtype],
+    per_token_quant: bool,
+    block_shape: Optional[list[int]],
+) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+    assert (quant_dtype == torch.float8_e4m3fn
+            or quant_dtype == torch.int8), "only fp8/int8 supported"
+
+    if block_shape is not None:
+        assert not per_token_quant
+        if quant_dtype == torch.int8:
+            w, w_s = per_block_cast_to_int8(w, block_shape)
+        else:
+            w, w_s = per_block_cast_to_fp8(w, block_shape)
+    else:
+        if quant_dtype == torch.int8:
+            w, w_s = ops.scaled_int8_quant(
+                w, w_s, use_per_token_if_dynamic=per_token_quant)
+        else:
+            w, w_s = ops.scaled_fp8_quant(
+                w, w_s, use_per_token_if_dynamic=per_token_quant)
+
+    return w, w_s
+
+
+def make_test_weight(
+    e: int,
+    rows: int,
+    cols: int,
+    in_dtype: torch.dtype = torch.bfloat16,
+    quant_dtype: Optional[torch.dtype] = None,
+    block_shape: Optional[list[int]] = None,
+    per_act_token_quant: bool = False,
+) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+    w_16 = torch.randn((e, rows, cols), device="cuda", dtype=in_dtype) / 15
+
+    if quant_dtype is not None:
+        w_l = [None] * e
+        w_s_l = [None] * e
+        for idx in range(e):
+            w_l[idx], w_s_l[idx] = moe_quantize_weights(
+                w_16[idx], None, quant_dtype, per_act_token_quant, block_shape)
+
+        w = torch.stack(w_l)
+        w_s = torch.stack(w_s_l)
+        if w_s.ndim == 2:
+            assert w_s.shape[-1] == 1
+            w_s = w_s.view(-1, 1, 1)
+
+        if block_shape is not None:
+            block_n, block_k = block_shape
+            n_tiles = (rows + block_n - 1) // block_n
+            k_tiles = (cols + block_k - 1) // block_k
+            assert w_s.shape == (e, n_tiles, k_tiles)
+    else:
+        w = w_16
+        w_s = None
+
+    return w_16, w, w_s
+
+
+def make_test_weights(
+    e: int,
+    n: int,
+    k: int,
+    in_dtype: torch.dtype = torch.bfloat16,
+    quant_dtype: Optional[torch.dtype] = None,
+    block_shape: Optional[list[int]] = None,
+    per_act_token_quant: bool = False,
+) -> tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor], torch.Tensor,
+           torch.Tensor, Optional[torch.Tensor]]:
+    return (
+        *make_test_weight(e, 2 * n, k, in_dtype, quant_dtype, block_shape,
+                          per_act_token_quant),
+        *make_test_weight(e, k, n, in_dtype, quant_dtype, block_shape,
+                          per_act_token_quant),
+    )

tests/kernels/quant_utils.py

@@ -5,7 +5,10 @@ from typing import Optional, Union
 
 import torch
 
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    group_broadcast)
 from vllm.platforms import current_platform
+from vllm.utils import round_up
 
 # Using the default value (240.0) from pytorch will cause accuracy
 # issue on dynamic quantization models. Here use 224.0 for rocm.
@@ -94,9 +97,15 @@ def ref_dynamic_per_tensor_fp8_quant(x: torch.tensor) \
     return ref_out, ref_scale.view((1, ))
 
 
-def native_w8a8_block_matmul(A: torch.Tensor, B: torch.Tensor,
-                             As: torch.Tensor, Bs: torch.Tensor, block_size,
-                             output_dtype):
+def native_w8a8_block_matmul(
+    A: torch.Tensor,
+    B: torch.Tensor,
+    As: torch.Tensor,
+    Bs: torch.Tensor,
+    block_size: list[int],
+    output_dtype: torch.dtype,
+    compute_type: torch.dtype = torch.float32,
+) -> torch.Tensor:
     """This function performs matrix multiplication with block-wise
     quantization using native torch.
     It is agnostic to the input data type and can be used for both int8 and
@@ -106,8 +115,8 @@ def native_w8a8_block_matmul(A: torch.Tensor, B: torch.Tensor,
     `Bs` (float32).
     The output is returned in the specified `output_dtype`.
     """
-    A = A.to(torch.float32)
-    B = B.to(torch.float32)
+    A = A.to(compute_type)
+    B = B.to(compute_type)
     assert A.shape[-1] == B.shape[-1]
     assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2
     assert len(block_size) == 2
@@ -122,11 +131,11 @@ def native_w8a8_block_matmul(A: torch.Tensor, B: torch.Tensor,
     As = As.reshape(M, As.shape[-1])
     n_tiles = (N + block_n - 1) // block_n
     k_tiles = (K + block_k - 1) // block_k
-    assert n_tiles == Bs.shape[0]
-    assert k_tiles == Bs.shape[1]
+    assert n_tiles == Bs.shape[0], f"{n_tiles} == {Bs.shape[0]}"
+    assert k_tiles == Bs.shape[1], f"{k_tiles} == {Bs.shape[1]}"
 
     C_shape = (M, N)
-    C = torch.zeros(C_shape, dtype=torch.float32, device=A.device)
+    C = torch.zeros(C_shape, dtype=compute_type, device=A.device)
 
     A_tiles = [
         A[:, i * block_k:min((i + 1) * block_k, K)] for i in range(k_tiles)
@@ -152,3 +161,152 @@ def native_w8a8_block_matmul(A: torch.Tensor, B: torch.Tensor,
 
     C = C.reshape(origin_C_shape).to(output_dtype)
     return C
+
+
+def native_per_token_group_quant_fp8(x,
+                                     group_size,
+                                     eps=1e-10,
+                                     dtype=torch.float8_e4m3fn):
+    """Function to perform per-token-group quantization on an input tensor
+    `x` using native torch."""
+    assert x.shape[-1] % group_size == 0, ("the last dimension of `x` must "
+                                           "be divisible by `group_size`")
+    assert x.is_contiguous(), "`x` is not contiguous"
+
+    finfo = torch.finfo(dtype)
+    fp8_min = finfo.min
+    fp8_max = finfo.max
+
+    x_ = x.reshape(x.numel() // group_size, group_size)
+    amax = x_.abs().max(dim=-1,
+                        keepdim=True)[0].clamp(min=eps).to(torch.float32)
+    x_s = amax / fp8_max
+    x_q = (x_ / x_s).clamp(min=fp8_min, max=fp8_max).to(dtype)
+    x_q = x_q.reshape(x.shape)
+    x_s = x_s.reshape(x.shape[:-1] + (x.shape[-1] // group_size, ))
+
+    return x_q, x_s
+
+
+def native_per_token_group_quant_int8(x,
+                                      group_size,
+                                      eps=1e-10,
+                                      dtype=torch.int8):
+    """Function to perform per-token-group quantization on an input tensor
+    `x` using native torch.
+
+    It converts the tensor values into int8 values and returns the
+    quantized tensor along with the scaling factor used for quantization.
+    """
+    assert (x.shape[-1] % group_size == 0
+            ), "the last dimension of `x` must be divisible by `group_size`"
+    assert x.is_contiguous(), "`x` is not contiguous"
+
+    iinfo = torch.iinfo(dtype)
+    int8_min = iinfo.min
+    int8_max = iinfo.max
+
+    x_ = x.reshape(x.numel() // group_size, group_size)
+    # Use float32 for scale calculation for stability
+    amax = x_.abs().max(dim=-1,
+                        keepdim=True)[0].clamp(min=eps).to(torch.float32)
+    x_s = amax / int8_max
+    x_q = (x_.to(torch.float32) / x_s).round().clamp(
+        min=int8_min, max=int8_max).to(dtype)  # Round before clamping
+    x_q = x_q.reshape(x.shape)
+    x_s = x_s.reshape(x.shape[:-1] + (x.shape[-1] // group_size, ))
+
+    return x_q, x_s
+
+
+DEFAULT_BLOCK_SHAPE = [128, 128]
+
+
+def per_block_cast_to_fp8(
+    x: torch.Tensor,
+    block_shape: list[int] = DEFAULT_BLOCK_SHAPE,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    block_m, block_n = block_shape
+    assert x.dim() == 2
+    m, n = x.shape
+    x_padded = torch.zeros((round_up(m, block_m), round_up(n, block_n)),
+                           dtype=x.dtype,
+                           device=x.device)
+    x_padded[:m, :n] = x
+    x_view = x_padded.view(-1, block_m, x_padded.size(1) // block_n, block_n)
+    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
+    x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
+    x_scaled_sub = x_scaled.view_as(x_padded)[:m, :n].contiguous()
+    scales = (x_amax / 448.0).view(x_view.size(0), x_view.size(2))
+    return x_scaled_sub, scales
+
+
+def per_block_cast_to_int8(
+    x: torch.Tensor,
+    block_shape: list[int] = DEFAULT_BLOCK_SHAPE,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    block_m, block_n = block_shape
+    assert x.dim() == 2
+    m, n = x.shape
+    x_padded = torch.zeros((round_up(m, block_m), round_up(n, block_n)),
+                           dtype=x.dtype,
+                           device=x.device)
+    x_padded[:m, :n] = x
+    x_view = x_padded.view(-1, block_m, x_padded.size(1) // block_n, block_n)
+    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
+    x_scaled = (x_view * (256.0 / x_amax)).to(torch.int8)
+    x_scaled_sub = x_scaled.view_as(x_padded)[:m, :n].contiguous()
+    scales = (x_amax / 256.0).view(x_view.size(0), x_view.size(2))
+    return x_scaled_sub, scales
+
+
+def dequant(
+    t: torch.Tensor,
+    scale: Optional[torch.Tensor],
+    block_shape: Optional[list[int]],
+    per_act_token_quant: bool,
+    out_dtype: Optional[torch.dtype] = torch.float32,
+) -> torch.Tensor:
+    if scale is not None:
+        f32 = torch.float32
+        if per_act_token_quant or block_shape is None:
+            return (t.to(f32) * scale).to(out_dtype)
+        else:
+            return (t.to(f32) * group_broadcast(scale, t.shape)).to(out_dtype)
+    else:
+        return t.to(out_dtype)
+
+
+def native_batched_masked_quant_matmul(
+    A: torch.Tensor,
+    B: torch.Tensor,
+    C: torch.Tensor,
+    num_expert_tokens: torch.Tensor,
+    A_scale: Optional[torch.Tensor] = None,
+    B_scale: Optional[torch.Tensor] = None,
+    block_shape: Optional[list[int]] = None,
+    per_act_token_quant: bool = False,
+) -> torch.Tensor:
+    num_expert_tokens_cpu = num_expert_tokens.clone()
+    num_expert_tokens_cpu = num_expert_tokens_cpu.to(device="cpu")
+    num_experts = num_expert_tokens.size(0)
+
+    for e in range(num_experts):
+        num_tokens = num_expert_tokens_cpu[e]
+        if A.dtype.itemsize == 1 and block_shape is not None:
+            assert A_scale is not None and B_scale is not None
+            tmp = native_w8a8_block_matmul(A[e], B[e], A_scale[e], B_scale[e],
+                                           block_shape, C.dtype)
+            C[e, :num_tokens, :] = tmp[:num_tokens, :]
+        elif A.dtype.itemsize == 1 and block_shape is None:
+            assert A_scale is not None and B_scale is not None
+            A_dq = dequant(A[e], A_scale[e], block_shape, per_act_token_quant)
+            B_dq = dequant(B[e], B_scale[e], block_shape, per_act_token_quant)
+            C[e, :num_tokens, :] = (
+                A_dq[:num_tokens] @ B_dq.transpose(0, 1)).to(C.dtype)
+        else:
+            assert A_scale is None
+            assert B_scale is None
+            C[e, :num_tokens, :] = A[e, :num_tokens, :] @ B[e].transpose(0, 1)
+
+    return C

tests/kernels/quantization/test_block_fp8.py

@@ -7,16 +7,10 @@ import itertools
 import pytest
 import torch
 
-from tests.kernels.quant_utils import native_w8a8_block_matmul
-from vllm.config import VllmConfig, set_current_vllm_config
-from vllm.model_executor.layers.activation import SiluAndMul
-from vllm.model_executor.layers.fused_moe import fused_moe
-from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
-    _valid_deep_gemm_shape, deep_gemm_moe_fp8)
-from vllm.model_executor.layers.fused_moe.fused_moe import (
-    fused_topk, modular_triton_fused_moe)
-from vllm.model_executor.layers.fused_moe.moe_align_block_size import (
-    moe_align_block_size)
+from tests.kernels.quant_utils import (native_per_token_group_quant_fp8,
+                                       native_w8a8_block_matmul,
+                                       per_block_cast_to_fp8)
+from vllm.config import VllmConfig
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8, w8a8_block_fp8_matmul)
 from vllm.platforms import current_platform
@@ -46,78 +40,10 @@ N = [128, 512, 7168, 7748, 13824]
 K = [256, 3884, 4096, 13824, 16384]
 # Deepseek-V3's intermediate size 18432, so N is 18432*2/8=4608 at TP8
 # and its hidden size is 7168.
-M_moe = [1, 2, 7, 83, 128, 2048, 1024 * 128]
-M_moe_dg = [128, 192, 1335, 2048]
-N_moe = [128, 256, 1024, 4608]  # [13824]
-K_moe = [256, 512, 7168]  # [13824]
 BLOCK_SIZE = [[128, 128]]
-E = [2, 8, 16, 24]  # [128, 256]
-TOP_KS = [1, 2, 6]
 OUT_DTYPES = [torch.bfloat16]  # [torch.float32, torch.half, torch.bfloat16]
 SEEDS = [0]
 
-
-def native_per_token_group_quant_fp8(x,
-                                     group_size,
-                                     eps=1e-10,
-                                     dtype=torch.float8_e4m3fn):
-    """Function to perform per-token-group quantization on an input tensor
-    `x` using native torch."""
-    assert x.shape[-1] % group_size == 0, ("the last dimension of `x` cannot "
-                                           "be divisible by `group_size`")
-    assert x.is_contiguous(), "`x` is not contiguous"
-
-    finfo = torch.finfo(dtype)
-    fp8_min = finfo.min
-    fp8_max = finfo.max
-
-    x_ = x.reshape(x.numel() // group_size, group_size)
-    amax = x_.abs().max(dim=-1,
-                        keepdim=True)[0].clamp(min=eps).to(torch.float32)
-    x_s = amax / fp8_max
-    x_q = (x_ / x_s).clamp(min=fp8_min, max=fp8_max).to(dtype)
-    x_q = x_q.reshape(x.shape)
-    x_s = x_s.reshape(x.shape[:-1] + (x.shape[-1] // group_size, ))
-
-    return x_q, x_s
-
-
-def torch_w8a8_block_fp8_moe(a, w1, w2, w1_s, w2_s, score, topk, block_shape):
-    """Fused moe with block-wise quantization using native torch."""
-    B, D = a.shape
-    a = a.view(B, -1, D).repeat(1, topk, 1).reshape(-1, D)
-    out = torch.zeros(B * topk, w2.shape[1], dtype=a.dtype, device=a.device)
-    score = torch.softmax(score, dim=-1, dtype=torch.float32)
-    topk_weight, topk_ids = torch.topk(score, topk)
-    topk_weight = topk_weight.view(-1)
-    topk_ids = topk_ids.view(-1)
-
-    _, block_k = block_shape[0], block_shape[1]
-    a_q, a_s = native_per_token_group_quant_fp8(a, block_k)
-    a_q = a_q.to(torch.float32)
-    for i in range(w1.shape[0]):
-        mask = topk_ids == i
-        if mask.sum():
-            inter_out = native_w8a8_block_matmul(a_q[mask],
-                                                 w1[i],
-                                                 a_s[mask],
-                                                 w1_s[i],
-                                                 block_shape,
-                                                 output_dtype=a.dtype)
-            act_out = SiluAndMul().forward_native(inter_out)
-            act_out_q, act_out_s = native_per_token_group_quant_fp8(
-                act_out, block_k)
-            act_out = act_out.to(torch.float32)
-            out[mask] = native_w8a8_block_matmul(act_out_q,
-                                                 w2[i],
-                                                 act_out_s,
-                                                 w2_s[i],
-                                                 block_shape,
-                                                 output_dtype=a.dtype)
-    return (out.view(B, -1, w2.shape[1]) *
-            topk_weight.view(B, -1, 1).to(out.dtype)).sum(dim=1)
-
-
 # Skip all tests if CUDA is not available
 pytest.importorskip("torch.cuda")
 
@@ -177,111 +103,6 @@ def test_w8a8_block_fp8_matmul(M, N, K, block_size, out_dtype, seed):
     assert rel_diff < 0.001
 
 
-@pytest.mark.parametrize(
-    "M,N,K,E,topk,block_size,dtype,seed",
-    itertools.product(M_moe, N_moe, K_moe, E, TOP_KS, BLOCK_SIZE, DTYPES,
-                      SEEDS))
-@torch.inference_mode()
-def test_w8a8_block_fp8_fused_moe(M, N, K, E, topk, block_size, dtype, seed):
-    if topk > E:
-        pytest.skip(f"Skipping test; topk={topk} > E={E}")
-
-    torch.manual_seed(seed)
-    factor_for_scale = 1e-2
-    fp8_info = torch.finfo(torch.float8_e4m3fn)
-    fp8_max, fp8_min = fp8_info.max, fp8_info.min
-
-    a = torch.randn((M, K), dtype=dtype) / 10
-
-    w1_bf16 = (torch.rand(
-        (E, 2 * N, K), dtype=torch.bfloat16) - 0.5) * 2 * fp8_max
-    w1 = w1_bf16.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
-    del w1_bf16
-
-    w2_bf16 = (torch.rand((E, K, N), dtype=torch.bfloat16) - 0.5) * 2 * fp8_max
-    w2 = w2_bf16.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
-    del w2_bf16
-
-    block_n, block_k = block_size[0], block_size[1]
-    n_tiles_w1 = (2 * N + block_n - 1) // block_n
-    n_tiles_w2 = (K + block_n - 1) // block_n
-    k_tiles_w1 = (K + block_k - 1) // block_k
-    k_tiles_w2 = (N + block_k - 1) // block_k
-
-    w1_s = torch.rand(
-        (E, n_tiles_w1, k_tiles_w1), dtype=torch.float32) * factor_for_scale
-    w2_s = torch.rand(
-        (E, n_tiles_w2, k_tiles_w2), dtype=torch.float32) * factor_for_scale
-
-    score = torch.randn((M, E), dtype=dtype)
-
-    m_fused_moe = modular_triton_fused_moe(use_fp8_w8a8=True,
-                                           use_int8_w8a8=False,
-                                           use_int8_w8a16=False,
-                                           use_int4_w4a16=False,
-                                           per_channel_quant=False,
-                                           block_shape=block_size)
-
-    # Set the context to avoid lots of warning spam.
-    with set_current_vllm_config(vllm_config):
-        out = fused_moe(
-            a,
-            w1,
-            w2,
-            score,
-            topk,
-            renormalize=False,
-            use_fp8_w8a8=True,
-            w1_scale=w1_s,
-            w2_scale=w2_s,
-            block_shape=block_size,
-        )
-        ref_out = torch_w8a8_block_fp8_moe(a, w1, w2, w1_s, w2_s, score, topk,
-                                           block_size)
-
-        topk_weights, topk_ids, _ = fused_topk(a, score, topk, False)
-        m_out = m_fused_moe(a,
-                            w1,
-                            w2,
-                            topk_weights,
-                            topk_ids,
-                            global_num_experts=E,
-                            w1_scale=w1_s,
-                            w2_scale=w2_s)
-
-    #print(f"{out.sum()=}")
-    #print(f"{ref_out.sum()=}")
-
-    rel_diff = (torch.mean(
-        torch.abs(out.to(torch.float32) - ref_out.to(torch.float32))) /
-                torch.mean(torch.abs(ref_out.to(torch.float32))))
-    assert rel_diff < 0.03
-
-    rel_diff = (torch.mean(
-        torch.abs(m_out.to(torch.float32) - ref_out.to(torch.float32))) /
-                torch.mean(torch.abs(ref_out.to(torch.float32))))
-    assert rel_diff < 0.03
-
-
-def per_block_cast_to_fp8(
-        x: torch.Tensor,
-        block_size_n: int = 128) -> tuple[torch.Tensor, torch.Tensor]:
-    assert x.dim() == 2
-    m, n = x.shape
-    x_padded = torch.zeros(
-        (deep_gemm.ceil_div(m, 128) * 128,
-         deep_gemm.ceil_div(n, block_size_n) * block_size_n),
-        dtype=x.dtype,
-        device=x.device)
-    x_padded[:m, :n] = x
-    x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, block_size_n)
-    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
-    x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
-    x_scaled_sub = x_scaled.view_as(x_padded)[:m, :n].contiguous()
-    scales = (x_amax / 448.0).view(x_view.size(0), x_view.size(2))
-    return x_scaled_sub, scales
-
-
 @pytest.mark.parametrize(
     "M,N,K,block_size,out_dtype,seed",
     itertools.product(M, N, K, BLOCK_SIZE, OUT_DTYPES, SEEDS))
@@ -324,187 +145,3 @@ def test_w8a8_block_fp8_deep_gemm_matmul(M, N, K, block_size, out_dtype, seed):
         torch.abs(out.to(torch.float32) - ref_out.to(torch.float32))) /
                 torch.mean(torch.abs(ref_out.to(torch.float32))))
     assert rel_diff < 0.001
-
-
-def fp8_perm(m, idx):
-    if torch.is_floating_point(m) and torch.finfo(m.dtype).bits == 8:
-        return m.view(dtype=torch.uint8)[idx, ...].view(dtype=m.dtype)
-    else:
-        return m[idx, ...]
-
-
-def _moe_permute(a, a_s, topk_ids, num_groups, topk, block_m):
-    M, K = a.shape
-
-    sorted_token_ids, m_indices, num_pad = moe_align_block_size(
-        topk_ids, block_m, num_groups, None, pad_sorted_ids=True)
-
-    num_tokens = topk * M
-
-    sorted_token_ids = sorted_token_ids.clamp(max=num_tokens - 1)
-    m_indices = torch.repeat_interleave(m_indices, block_m, dim=0)
-    inv_perm = torch.argsort(sorted_token_ids)[:M * topk]
-
-    a = fp8_perm(a, sorted_token_ids // topk)
-    if a_s is not None:
-        a_s = a_s[sorted_token_ids // topk]
-
-    return a, a_s, m_indices, inv_perm
-
-
-def _moe_unpermute(out, inv_perm, topk, K, topk_weight):
-    M = topk_weight.shape[0]
-    out = out[inv_perm, ...]
-    tmp_out = out.view(-1, topk, K)
-    return (tmp_out * topk_weight.view(M, -1, 1).to(out.dtype)).sum(dim=1)
-
-
-def deep_gemm_w8a8_block_fp8_moe(M, K, a, w1, w2, w1_s, w2_s, score, topk,
-                                 block_shape):
-    """Fused moe with block-wise quantization using DeepGemm grouped gemm."""
-    num_groups = w1.shape[0]
-    M, K = a.shape
-    N = w2.shape[-1]
-
-    topk_weight, topk_ids, token_expert_indices = fused_topk(
-        a, score.float(), topk, False)
-
-    block_m = deep_gemm.get_m_alignment_for_contiguous_layout()
-
-    _, block_k = block_shape[0], block_shape[1]
-
-    a_q, a_s = per_token_group_quant_fp8(a, block_m)
-
-    a_q, a_s, m_indices, inv_perm = _moe_permute(a_q, a_s, topk_ids,
-                                                 num_groups, topk, block_m)
-
-    inter_out = torch.zeros((a_q.shape[0], N * 2),
-                            dtype=torch.bfloat16,
-                            device=a.device)
-
-    deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_contiguous((a_q, a_s), (w1, w1_s),
-                                                        inter_out, m_indices)
-
-    act_out = SiluAndMul().forward_native(inter_out)
-    act_out_q, act_out_s = per_token_group_quant_fp8(act_out, block_k)
-
-    out = torch.zeros(a_q.shape[0], K, dtype=torch.bfloat16, device=a.device)
-
-    deep_gemm.m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(
-        (act_out_q, act_out_s), (w2, w2_s), out, m_indices)
-
-    final_out = _moe_unpermute(out, inv_perm, topk, K, topk_weight)
-
-    return final_out
-
-
-@pytest.mark.parametrize(
-    "M,N,K,E,topk,seed",
-    itertools.product(M_moe_dg, N_moe, K_moe, E, TOP_KS, SEEDS))
-@pytest.mark.skipif(not dg_available, reason="DeepGemm kernels not available.")
-@torch.inference_mode()
-def test_w8a8_block_fp8_deep_gemm_fused_moe(M, N, K, E, topk, seed,
-                                            monkeypatch):
-    if topk > E:
-        pytest.skip(f"Skipping test: topk={topk} > E={E}")
-
-    if not _valid_deep_gemm_shape(M, N, K):
-        pytest.skip(f"Skipping test: invalid size m={M}, n={N}, k={K}")
-
-    chunk_size = 1024
-
-    torch.manual_seed(seed)
-
-    monkeypatch.setenv("VLLM_FUSED_MOE_CHUNK_SIZE", str(chunk_size))
-
-    block_m = deep_gemm.get_m_alignment_for_contiguous_layout()
-    block_size = [block_m, block_m]
-    dtype = torch.bfloat16
-
-    fp8_info = torch.finfo(torch.float8_e4m3fn)
-    fp8_max, fp8_min = fp8_info.max, fp8_info.min
-
-    a = torch.randn((M, K), dtype=dtype) / 10
-
-    w1_bf16 = ((torch.rand((E, 2 * N, K), dtype=torch.bfloat16) - 0.5) * 2 *
-               fp8_max).clamp(min=fp8_min, max=fp8_max)
-
-    w2_bf16 = ((torch.rand((E, K, N), dtype=torch.bfloat16) - 0.5) * 2 *
-               fp8_max).clamp(min=fp8_min, max=fp8_max)
-
-    score = torch.randn((M, E), dtype=dtype)
-
-    block_n, block_k = block_size[0], block_size[1]
-    n_tiles_w1 = ((2 * N) + block_n - 1) // block_n
-    k_tiles_w1 = (K + block_k - 1) // block_k
-    n_tiles_w2 = (K + block_n - 1) // block_n
-    k_tiles_w2 = (N + block_k - 1) // block_k
-
-    w1 = torch.empty_like(w1_bf16, dtype=torch.float8_e4m3fn)
-    w2 = torch.empty_like(w2_bf16, dtype=torch.float8_e4m3fn)
-
-    w1_s = torch.empty((E, n_tiles_w1, k_tiles_w1), dtype=torch.float32)
-    w2_s = torch.empty((E, n_tiles_w2, k_tiles_w2), dtype=torch.float32)
-
-    w1_s = deep_gemm.get_col_major_tma_aligned_tensor(w1_s).contiguous()
-    w2_s = deep_gemm.get_col_major_tma_aligned_tensor(w2_s).contiguous()
-
-    assert w1_s.shape == (E, (2 * N + 127) // 128, (K + 127) // 128)
-    assert (w2.shape[-2] + block_n - 1) // block_n == w2_s.shape[-2]
-
-    for i in range(E):
-        w1[i], w1_s[i] = per_block_cast_to_fp8(w1_bf16[i])
-        w2[i], w2_s[i] = per_block_cast_to_fp8(w2_bf16[i])
-
-    # Note: for now use_compile will error out if the problem size is
-    # large enough to trigger chunking. I'm leaving the flag and
-    # setup code in case we are able to revisit this later.
-    use_compile = False
-
-    use_cudagraph = (chunk_size < M and N >= 1024 and K >= 1024
-                     and current_platform.is_cuda_alike())
-
-    # Set the context to avoid lots of warning spam.
-    with set_current_vllm_config(vllm_config):
-        if M >= 128:
-            ref_out = deep_gemm_w8a8_block_fp8_moe(M, K, a, w1, w2, w1_s, w2_s,
-                                                   score, topk, block_size)
-        else:
-            ref_out = torch_w8a8_block_fp8_moe(a, w1, w2, w1_s, w2_s, score,
-                                               topk, block_size)
-
-        topk_weights, topk_ids, token_expert_indices = fused_topk(
-            a, score.float(), topk, False)
-
-        if use_compile:
-            deep_gemm_moe_fp8_fn = torch.compile(deep_gemm_moe_fp8,
-                                                 backend="inductor",
-                                                 fullgraph=True)
-            torch._dynamo.mark_dynamic(a, 0)
-            torch._dynamo.mark_dynamic(topk_weights, 0)
-            torch._dynamo.mark_dynamic(topk_ids, 0)
-        else:
-            deep_gemm_moe_fp8_fn = deep_gemm_moe_fp8
-
-        out = deep_gemm_moe_fp8_fn(a, w1, w2, w1_s, w2_s, topk_weights,
-                                   topk_ids)
-
-        if use_cudagraph:
-            out.fill_(0)
-            stream = torch.cuda.Stream()
-            graph = torch.cuda.CUDAGraph()
-            with torch.cuda.graph(graph, stream=stream):
-                out = deep_gemm_moe_fp8_fn(a, w1, w2, w1_s, w2_s, topk_weights,
-                                           topk_ids)
-            torch.cuda.synchronize()
-            graph.replay()
-            torch.cuda.synchronize()
-
-    #print(f"{out.sum()=}")
-    #print(f"{ref_out.sum()=}")
-
-    rel_diff = (torch.mean(
-        torch.abs(out.to(torch.float32) - ref_out.to(torch.float32))) /
-                torch.mean(torch.abs(ref_out.to(torch.float32))))
-
-    assert rel_diff < 0.03

tests/kernels/quantization/test_block_int8.py

@@ -8,9 +8,7 @@ import pytest
 import torch
 
 from tests.kernels.quant_utils import native_w8a8_block_matmul
-from vllm.config import VllmConfig, set_current_vllm_config
-from vllm.model_executor.layers.activation import SiluAndMul
-from vllm.model_executor.layers.fused_moe import fused_moe
+from vllm.config import VllmConfig
 from vllm.model_executor.layers.quantization.utils.int8_utils import (
     w8a8_block_int8_matmul)
 from vllm.platforms import current_platform
@@ -23,82 +21,10 @@ vllm_config = VllmConfig()
 vllm_config.scheduler_config.max_num_seqs = 128
 vllm_config.scheduler_config.max_model_len = 8192
 
-
-# For test
-def native_per_token_group_quant_int8(x,
-                                      group_size,
-                                      eps=1e-10,
-                                      dtype=torch.int8):
-    """Function to perform per-token-group quantization on an input tensor
-    `x` using native torch.
-
-    It converts the tensor values into int8 values and returns the
-    quantized tensor along with the scaling factor used for quantization.
-    """
-    assert (x.shape[-1] % group_size == 0
-            ), "the last dimension of `x` cannot be divisible by `group_size`"
-    assert x.is_contiguous(), "`x` is not contiguous"
-
-    iinfo = torch.iinfo(dtype)
-    int8_min = iinfo.min
-    int8_max = iinfo.max
-
-    x_ = x.reshape(x.numel() // group_size, group_size)
-    # Use float32 for scale calculation for stability
-    amax = x_.abs().max(dim=-1,
-                        keepdim=True)[0].clamp(min=eps).to(torch.float32)
-    x_s = amax / int8_max
-    x_q = (x_.to(torch.float32) / x_s).round().clamp(
-        min=int8_min, max=int8_max).to(dtype)  # Round before clamping
-    x_q = x_q.reshape(x.shape)
-    x_s = x_s.reshape(x.shape[:-1] + (x.shape[-1] // group_size, ))
-
-    return x_q, x_s
-
-
-# For test
-def torch_w8a8_block_int8_moe(a, w1, w2, w1_s, w2_s, score, topk, block_shape):
-    """This function performs fused moe with block-wise quantization using
-    native torch."""
-    B, D = a.shape
-    a = a.view(B, -1, D).repeat(1, topk, 1).reshape(-1, D)
-    out = torch.zeros(B * topk, w2.shape[1], dtype=a.dtype, device=a.device)
-    score = torch.softmax(score, dim=-1, dtype=torch.float32)
-    topk_weight, topk_ids = torch.topk(score, topk)
-    topk_weight = topk_weight.view(-1)
-    topk_ids = topk_ids.view(-1)
-
-    _, block_k = block_shape[0], block_shape[1]
-    a_q, a_s = native_per_token_group_quant_int8(a, block_k)
-    for i in range(w1.shape[0]):
-        mask = topk_ids == i
-        if mask.sum():
-            inter_out = native_w8a8_block_matmul(a_q[mask],
-                                                 w1[i],
-                                                 a_s[mask],
-                                                 w1_s[i],
-                                                 block_shape,
-                                                 output_dtype=a.dtype)
-            act_out = SiluAndMul().forward_native(inter_out)
-            act_out_q, act_out_s = native_per_token_group_quant_int8(
-                act_out, block_k)
-            act_out = act_out.to(torch.float32)
-            out[mask] = native_w8a8_block_matmul(act_out_q,
-                                                 w2[i],
-                                                 act_out_s,
-                                                 w2_s[i],
-                                                 block_shape,
-                                                 output_dtype=a.dtype)
-    return (out.view(B, -1, w2.shape[1]) *
-            topk_weight.view(B, -1, 1).to(out.dtype)).sum(dim=1)
-
-
 DTYPES = [torch.half, torch.bfloat16]
 M = [1, 33, 64, 222]
 N = [128, 1024]
 K = [256, 4096]
-E = [8, 24]
-TOP_KS = [2, 6]
 # BLOCK_SIZE = [[64, 64], [64, 128], [128, 64], [128, 128]]
 BLOCK_SIZE = [[128, 128]]
 SEEDS = [0]
@@ -140,63 +66,3 @@ def test_w8a8_block_int8_matmul(M, N, K, block_size, out_dtype, seed):
         torch.abs(out.to(torch.float32) - ref_out.to(torch.float32))) /
                 torch.mean(torch.abs(ref_out.to(torch.float32))))
     assert rel_diff < 0.001
-
-
-@pytest.mark.parametrize(
-    "M, N, K, E, topk, block_size, dtype, seed",
-    itertools.product(M, N, K, E, TOP_KS, BLOCK_SIZE, DTYPES, SEEDS))
-@torch.inference_mode()
-def test_w8a8_block_int8_fused_moe(M, N, K, E, topk, block_size, dtype, seed):
-    """Tests the fused_moe kernel with W8A8 INT8 block quantization against a
-    native torch reference."""
-    torch.manual_seed(seed)
-    # Use a smaller factor for scale initialization to prevent large
-    # values/overflow especially when output dtype might be float16
-    factor_for_scale = 1e-2
-    int8_info = torch.iinfo(torch.int8)
-    int8_max, int8_min = int8_info.max, int8_info.min
-
-    a = torch.randn((M, K), dtype=dtype) / 10
-
-    w1_fp32 = (torch.rand(
-        (E, 2 * N, K), dtype=torch.float32) - 0.5) * 2 * int8_max
-    w1 = w1_fp32.clamp(min=int8_min, max=int8_max).to(torch.int8)
-
-    w2_fp32 = (torch.rand((E, K, N), dtype=torch.float32) - 0.5) * 2 * int8_max
-    w2 = w2_fp32.clamp(min=int8_min, max=int8_max).to(torch.int8)
-
-    block_n, block_k = block_size[0], block_size[1]
-    n_tiles_w1 = (2 * N + block_n - 1) // block_n
-    n_tiles_w2 = (K + block_n - 1) // block_n
-    k_tiles_w1 = (K + block_k - 1) // block_k
-    k_tiles_w2 = (N + block_k - 1) // block_k
-
-    w1_s = (torch.rand(
-        (E, n_tiles_w1, k_tiles_w1), dtype=torch.float32) * factor_for_scale)
-    w2_s = (torch.rand(
-        (E, n_tiles_w2, k_tiles_w2), dtype=torch.float32) * factor_for_scale)
-
-    score = torch.randn((M, E), dtype=dtype)
-
-    # Set the context to avoid lots of warning spam.
-    with set_current_vllm_config(vllm_config):
-        out = fused_moe(
-            a,
-            w1,
-            w2,
-            score,
-            topk,
-            renormalize=False,
-            use_int8_w8a8=True,
-            w1_scale=w1_s,
-            w2_scale=w2_s,
-            block_shape=block_size,
-        )
-        ref_out = torch_w8a8_block_int8_moe(a, w1, w2, w1_s, w2_s, score, topk,
-                                            block_size)
-
-    # Check results
-    rel_diff = (torch.mean(
-        torch.abs(out.to(torch.float32) - ref_out.to(torch.float32))) /
-                torch.mean(torch.abs(ref_out.to(torch.float32))))
-    assert rel_diff < 0.06

tests/kernels/quantization/test_machete_mm.py

@@ -14,6 +14,8 @@ import torch
 
 from tests.kernels.utils import opcheck
 from vllm import _custom_ops as ops
+from vllm.model_executor.layers.quantization.utils.machete_utils import (
+    query_machete_supported_group_sizes)
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
     pack_rows, quantize_weights)
 from vllm.platforms import current_platform
@@ -46,8 +48,6 @@ MNK_SHAPES = [
     (1024, 8192, 4096),
 ]
 
-GROUP_SIZES_TO_TEST: list[Optional[int]] = [128, -1]
-
 
 @dataclass
 class TypeConfig:
@@ -139,7 +139,7 @@ def maybe_convert_zeropoints(zps: Optional[torch.Tensor], s: torch.Tensor):
 
 def group_size_valid(shape: tuple[int, int, int],
                      group_size: Optional[int]) -> bool:
-    return group_size is None or group_size == -1 or group_size % shape[2] == 0
+    return group_size is None or group_size == -1 or shape[2] % group_size == 0
 
 
 def machete_quantize_and_pack(atype: torch.dtype,
@@ -270,7 +270,7 @@ def test_machete_all_schedules(shape, types: TypeConfig):
     if types.group_scale_type is None:
         group_sizes = [None]
     else:
-        group_sizes = GROUP_SIZES_TO_TEST
+        group_sizes = query_machete_supported_group_sizes(types.act_type)
 
     for group_size in group_sizes:
         if not group_size_valid(shape, group_size):
@@ -299,7 +299,7 @@ def test_machete_heuristic(shape, types: TypeConfig):
     if types.group_scale_type is None:
         group_sizes = [None]
     else:
-        group_sizes = GROUP_SIZES_TO_TEST
+        group_sizes = query_machete_supported_group_sizes(types.act_type)
 
     for group_size in group_sizes:
         if not group_size_valid(shape, group_size):

tests/kernels/utils.py

@@ -13,8 +13,11 @@ import pytest
 import torch
 from torch._prims_common import TensorLikeType
 
+from tests.kernels.quant_utils import native_w8a8_block_matmul
 from vllm.attention import AttentionBackend, AttentionMetadata, AttentionType
 from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.fused_moe.utils import (
+    moe_kernel_quantize_input)
 from vllm.platforms.interface import _Backend
 from vllm.utils import (STR_BACKEND_ENV_VAR, STR_FLASH_ATTN_VAL,
                         STR_XFORMERS_ATTN_VAL, make_tensor_with_pad)
@@ -1054,32 +1057,77 @@ def compute_max_diff(output, output_ref):
         torch.abs(output_ref))
 
 
-def torch_experts(a: torch.Tensor,
-                  w1: torch.Tensor,
-                  w2: torch.Tensor,
-                  topk_weight: torch.Tensor,
-                  topk_ids: torch.Tensor,
-                  global_num_experts: int = -1,
-                  expert_map: Optional[torch.Tensor] = None) -> torch.Tensor:
+def torch_experts(
+    a: torch.Tensor,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    topk_weight: torch.Tensor,
+    topk_ids: torch.Tensor,
+    global_num_experts: int = -1,
+    expert_map: Optional[torch.Tensor] = None,
+    w1_scale: Optional[torch.Tensor] = None,
+    w2_scale: Optional[torch.Tensor] = None,
+    a1_scale: Optional[torch.Tensor] = None,
+    a2_scale: Optional[torch.Tensor] = None,
+    quant_dtype: Optional[torch.dtype] = None,
+    per_act_token_quant=False,
+    block_shape: Optional[list[int]] = None,
+) -> torch.Tensor:
     assert (global_num_experts == -1
             or (global_num_experts == w1.shape[0] and expert_map is None)
             or (expert_map is not None
                 and global_num_experts == expert_map.shape[0]))
+
+    M, K = a.shape
     topk = topk_ids.shape[1]
-    B, D = a.shape
-    a = a.view(B, -1, D).repeat(1, topk, 1).reshape(-1, D)
-    out = torch.zeros(B * topk, w2.shape[1], dtype=a.dtype, device=a.device)
-    topk_weight = topk_weight.view(-1)
+
+    a = a.view(M, -1, K).repeat(1, topk, 1).reshape(-1, K)
+
+    out = torch.zeros(M * topk, w2.shape[1], dtype=a.dtype, device=a.device)
+
+    a, a_scale = moe_kernel_quantize_input(a, None, quant_dtype,
+                                           per_act_token_quant, block_shape)
+
+    num_experts = w1.shape[0]
+
     topk_ids = topk_ids.view(-1)
     if expert_map is not None:
         topk_ids = expert_map[topk_ids]
-    for i in range(w1.shape[0]):
+
+    for i in range(num_experts):
         mask = topk_ids == i
         if mask.sum():
-            out[mask] = SiluAndMul()(
-                a[mask] @ w1[i].transpose(0, 1)) @ w2[i].transpose(0, 1)
-    return (out.view(B, -1, w2.shape[1]) *
-            topk_weight.view(B, -1, 1).to(out.dtype)).sum(dim=1)
+            if quant_dtype is None:
+                tmp1 = a[mask] @ w1[i].transpose(0, 1)
+                tmp2 = SiluAndMul()(tmp1)
+                out[mask] = tmp2 @ w2[i].transpose(0, 1)
+            elif block_shape is not None:
+                assert (a_scale is not None and w1_scale is not None
+                        and w2_scale is not None)
+                tmp1 = native_w8a8_block_matmul(a[mask], w1[i], a_scale[mask],
+                                                w1_scale[i], block_shape,
+                                                out.dtype)
+                tmp2 = SiluAndMul()(tmp1)
+                tmp2, b_scale = moe_kernel_quantize_input(
+                    tmp2, None, quant_dtype, per_act_token_quant, block_shape)
+
+                out[mask] = native_w8a8_block_matmul(tmp2, w2[i], b_scale,
+                                                     w2_scale[i], block_shape,
+                                                     out.dtype)
+            else:
+                assert (a_scale is not None and w1_scale is not None
+                        and w2_scale is not None)
+                f32 = torch.float32
+                scales = a_scale if a_scale.numel() == 1 else a_scale[mask]
+                tmp1 = a[mask].to(f32) * scales
+                w1_dq = (w1[i].to(f32) * w1_scale[i]).transpose(0, 1)
+                tmp1 = tmp1 @ w1_dq
+                tmp2 = SiluAndMul()(tmp1)
+                w2_dq = (w2[i].to(f32) * w2_scale[i]).transpose(0, 1)
+                out[mask] = (tmp2 @ w2_dq).to(out.dtype)
+
+    return (out.view(M, -1, w2.shape[1]) *
+            topk_weight.view(M, -1, 1).to(out.dtype)).sum(dim=1)
 
 
 def torch_moe(a: torch.Tensor,

tests/models/language/generation/test_common.py

@@ -78,7 +78,7 @@ AITER_MODEL_LIST = [
         ),
         pytest.param(
             "Qwen/Qwen2.5-0.5B-Instruct",  # qwen2
-            marks=[pytest.mark.core_model],
+            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
         ),
         pytest.param(
             "Qwen/Qwen3-8B",  # qwen (text-only)
@@ -87,6 +87,7 @@ AITER_MODEL_LIST = [
         pytest.param("bigcode/starcoder2-3b"),  # starcoder2
         pytest.param(
             "TitanML/tiny-mixtral",  # mixtral
+            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
         )
     ])
 @pytest.mark.parametrize("max_tokens", [32])

tests/models/language/generation/test_granitemoehybrid.py

@@ -1,42 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import pytest
-
-from ...utils import check_logprobs_close
-
-# Path of the checkpoints
-MODELS = [
-    "ibm-granite/granite-4.0-tiny-preview",
-]
-
-
-@pytest.mark.skip(
-    reason="Granite 4.0 is not yet available in huggingface transformers")
-@pytest.mark.parametrize("model", MODELS)
-@pytest.mark.parametrize("dtype", ["float16", "bfloat16"])
-@pytest.mark.parametrize("max_tokens", [64])
-@pytest.mark.parametrize("num_logprobs", [5])
-def test_model_equivalence_to_hf_greedy(
-    hf_runner,
-    vllm_runner,
-    example_prompts,
-    model: str,
-    dtype: str,
-    max_tokens: int,
-    num_logprobs: int,
-):
-    with vllm_runner(model, dtype=dtype) as vllm_model:
-        vllm_outputs = vllm_model.generate_greedy_logprobs(
-            example_prompts, max_tokens, num_logprobs)
-
-    with hf_runner(model, dtype=dtype) as hf_model:
-        hf_outputs = hf_model.generate_greedy_logprobs_limit(
-            example_prompts, max_tokens, num_logprobs)
-
-    check_logprobs_close(
-        outputs_0_lst=hf_outputs,
-        outputs_1_lst=vllm_outputs,
-        name_0="hf",
-        name_1="vllm",
-    )

tests/models/language/generation/test_hybrid.py

@@ -9,6 +9,9 @@ from vllm.sampling_params import SamplingParams
 
 from ...utils import check_logprobs_close, check_outputs_equal
 
+# Mark all tests as hybrid
+pytestmark = pytest.mark.hybrid_model
+
 # NOTE: The first model in each list is taken as the primary model,
 # meaning that it will be used in all tests in this file
 # The rest of the models will only be tested by test_models
@@ -25,8 +28,9 @@ SSM_MODELS = [
 
 HYBRID_MODELS = [
     "ai21labs/Jamba-tiny-dev",
-    # NOTE: ibm-granite/granite-4.0-tiny-preview are skipped currently as
-    # it is not yet available in huggingface transformers
+    # NOTE: Currently the test failes due to HF transformers issue fixed in:
+    # https://github.com/huggingface/transformers/pull/39033
+    # We will enable vLLM test for Granite after next HF transformers release.
     # "ibm-granite/granite-4.0-tiny-preview",
     # NOTE: Running Plamo2 in transformers implementation requires to install
     # causal-conv1d package, which is not listed as a test dependency as it's

tests/models/multimodal/generation/test_common.py

@@ -309,6 +309,34 @@ VLM_TEST_SETTINGS = {
         num_logprobs=10,
         marks=[large_gpu_mark(min_gb=32)],
     ),
+    "glm4_1v": VLMTestInfo(
+        models=["THUDM/GLM-4.1V-9B-Thinking"],
+        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
+        prompt_formatter=lambda img_prompt: f"<|user|>\n{img_prompt}<|assistant|>",  # noqa: E501
+        img_idx_to_prompt=lambda idx: "<|begin_of_image|><|image|><|end_of_image|>", # noqa: E501
+        video_idx_to_prompt=lambda idx: "<|begin_of_video|><|video|><|end_of_video|>", # noqa: E501
+        max_model_len=2048,
+        max_num_seqs=2,
+        get_stop_token_ids=lambda tok: [151329, 151336, 151338],
+        num_logprobs=10,
+        image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
+        auto_cls=AutoModelForImageTextToText,
+    ),
+    "glm4_1v-video": VLMTestInfo(
+        models=["THUDM/GLM-4.1V-9B-Thinking"],
+        # GLM4.1V require include video metadata for input
+        test_type=VLMTestType.CUSTOM_INPUTS,
+        max_model_len=4096,
+        max_num_seqs=2,
+        auto_cls=AutoModelForImageTextToText,
+        patch_hf_runner=model_utils.glm4_1v_patch_hf_runner,
+        custom_test_opts=[CustomTestOptions(
+            inputs=custom_inputs.video_with_metadata_glm4_1v(),
+            limit_mm_per_prompt={"video": 1},
+        )],
+        # This is needed to run on machine with 24GB VRAM
+        vllm_runner_kwargs={"gpu_memory_utilization": 0.95},
+    ),
     "h2ovl": VLMTestInfo(
         models = [
             "h2oai/h2ovl-mississippi-800m",

tests/models/multimodal/generation/vlm_utils/custom_inputs.py

@@ -129,3 +129,23 @@ def windows_attention_image_qwen2_5_vl():
 
     wrapped_sf = ImageSizeWrapper(type=SizeType.SIZE_FACTOR, data=[0.5])
     return build_single_image_inputs([image], [prompt], wrapped_sf)
+
+
+def video_with_metadata_glm4_1v():
+    video_array = VIDEO_ASSETS[0].np_ndarrays
+    metadata = VIDEO_ASSETS[0].metadata
+    question = "Describe the video."
+    video_prompt = "<|begin_of_video|><|video|><|end_of_video|>"
+    formatted_prompt = f"<|user|>\n{video_prompt}{question}<|assistant|>\n"
+
+    scales = [0.1, 0.2, 0.25]
+    video_input = [[(rescale_video_size(video_array, scale), metadata)]
+                   for scale in scales]
+    prompts = [formatted_prompt] * len(video_input)
+
+    return [
+        PromptWithMultiModalInput(
+            prompts=prompts,
+            video_data=video_input,
+        )
+    ]