[CPU] Refactor CPU W8A8 scaled_mm (#23071)

Signed-off-by: jiang1.li <jiang1.li@intel.com>

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

@@ -46,6 +46,11 @@ function cpu_tests() {
     set -e
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
 
+  # Run kernel tests
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
+    set -e
+    pytest -v -s tests/kernels/test_onednn.py"
+
   # Run basic model test
   docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
@@ -99,4 +104,4 @@ function cpu_tests() {
 
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 1.5h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 2h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"

cmake/cpu_extension.cmake

@@ -182,17 +182,17 @@ endif()
 #
 # Build oneDNN for W8A8 GEMM kernels (only for x86-AVX512 /ARM platforms)
 # Flag to enable ACL kernels for AARCH64 platforms
-if ( VLLM_BUILD_ACL STREQUAL "ON")
+if (VLLM_BUILD_ACL STREQUAL "ON")
     set(USE_ACL ON)
 else()
     set(USE_ACL OFF)
 endif()
 
-if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR ASIMD_FOUND)
+if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR ASIMD_FOUND OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
     FetchContent_Declare(
         oneDNN
         GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-        GIT_TAG  v3.8.1
+        GIT_TAG v3.9
         GIT_PROGRESS TRUE
         GIT_SHALLOW TRUE
     )
@@ -217,38 +217,23 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR ASIMD_FOUND)
     set(ONEDNN_ENABLE_ITT_TASKS "OFF")
     set(ONEDNN_ENABLE_MAX_CPU_ISA "OFF")
     set(ONEDNN_ENABLE_CPU_ISA_HINTS "OFF")
+    set(ONEDNN_VERBOSE "OFF")
     set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
 
     FetchContent_MakeAvailable(oneDNN)
-    
-    list(APPEND LIBS dnnl)
-elseif(POWER10_FOUND)
-    FetchContent_Declare(
-        oneDNN
-        GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-        GIT_TAG v3.7.2
-        GIT_PROGRESS TRUE
-        GIT_SHALLOW TRUE
+    add_library(dnnl_ext OBJECT "csrc/cpu/dnnl_helper.cpp")
+    target_include_directories(
+        dnnl_ext
+        PUBLIC ${oneDNN_SOURCE_DIR}/include
+        PUBLIC ${oneDNN_BINARY_DIR}/include
+        PRIVATE ${oneDNN_SOURCE_DIR}/src
     )
-
-    set(ONEDNN_LIBRARY_TYPE "STATIC")
-    set(ONEDNN_BUILD_DOC "OFF")
-    set(ONEDNN_BUILD_EXAMPLES "OFF")
-    set(ONEDNN_BUILD_TESTS "OFF")
-    set(ONEDNN_ENABLE_WORKLOAD "INFERENCE")
-    set(ONEDNN_ENABLE_PRIMITIVE "MATMUL;REORDER")
-    set(ONEDNN_BUILD_GRAPH "OFF")
-    set(ONEDNN_ENABLE_JIT_PROFILING "OFF")
-    set(ONEDNN_ENABLE_ITT_TASKS "OFF")
-    set(ONEDNN_ENABLE_MAX_CPU_ISA "OFF")
-    set(ONEDNN_ENABLE_CPU_ISA_HINTS "OFF")
-    set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
-
-    set(DNNL_CPU_RUNTIME "OMP")
-
-    FetchContent_MakeAvailable(oneDNN)
-
-    list(APPEND LIBS dnnl)
+    target_link_libraries(dnnl_ext dnnl)
+    target_compile_options(dnnl_ext PRIVATE ${CXX_COMPILE_FLAGS} -fPIC)
+    list(APPEND LIBS dnnl_ext)
+    set(USE_ONEDNN ON)
+else()
+    set(USE_ONEDNN OFF)
 endif()
 
 message(STATUS "CPU extension compile flags: ${CXX_COMPILE_FLAGS}")
@@ -275,7 +260,6 @@ set(VLLM_EXT_SRC
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     set(VLLM_EXT_SRC
-        "csrc/cpu/quant.cpp"
         "csrc/cpu/shm.cpp"
         ${VLLM_EXT_SRC})
     if (ENABLE_AVX512BF16 AND ENABLE_AVX512VNNI)
@@ -289,14 +273,11 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
             ${VLLM_EXT_SRC})
         add_compile_definitions(-DCPU_CAPABILITY_AVX512)
     endif()
-elseif(POWER10_FOUND)
-    set(VLLM_EXT_SRC
-        "csrc/cpu/quant.cpp"
-        ${VLLM_EXT_SRC})
 endif()
-if (ASIMD_FOUND)
+
+if(USE_ONEDNN)
     set(VLLM_EXT_SRC
-        "csrc/cpu/quant.cpp"
+        "csrc/cpu/dnnl_kernels.cpp"
         ${VLLM_EXT_SRC})
 endif()
 

csrc/cpu/cpu_types_x86.hpp

@@ -89,7 +89,7 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
 
   explicit FP16Vec16(const FP32Vec16&);
 
-  void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
+  void save(void* ptr) const { _mm256_storeu_si256((__m256i*)ptr, reg); }
 
   void save(void* ptr, const int elem_num) const {
     constexpr uint32_t M = 0xFFFFFFFF;
@@ -126,7 +126,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
 
   explicit BF16Vec16(const FP32Vec16&);
 
-  void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
+  void save(void* ptr) const { _mm256_storeu_si256((__m256i*)ptr, reg); }
 
   void save(void* ptr, const int elem_num) const {
     constexpr uint32_t M = 0xFFFFFFFF;
@@ -180,8 +180,8 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
             (__m128i)vec8_data.reg, 1)) {}
 
   void save(void* ptr) const {
-    *reinterpret_cast<__m256i*>(ptr) = reg_low;
-    *reinterpret_cast<__m256i*>((__m256i*)ptr + 1) = reg_high;
+    _mm256_storeu_si256((__m256i*)ptr, reg_low);
+    _mm256_storeu_si256((__m256i*)ptr + 1, reg_high);
   }
 };
 #endif

csrc/cpu/dnnl_helper.cpp

+#include <list>
+#include <optional>
+
+#include "common/memory_desc.hpp"
+#include "common/memory.hpp"
+
+#include "dnnl_helper.h"
+
+static dnnl::engine& default_engine() {
+  static dnnl::engine engine(dnnl::engine::kind::cpu, 0);
+  return engine;
+}
+
+static dnnl::stream& default_stream() {
+  static dnnl::stream stream(default_engine());
+  return stream;
+}
+
+void release_dnnl_matmul_handler(int64_t handler) {
+  DNNLMatMulPrimitiveHandler* ptr =
+      reinterpret_cast<DNNLMatMulPrimitiveHandler*>(handler);
+  delete ptr;
+}
+
+template <typename KT, typename VT>
+class DNNLPrimitiveCache {
+ public:
+  using cache_value_t = std::pair<KT, VT>;
+  using result_value_t = VT;
+  using container_t = std::list<cache_value_t>;
+  using value_iterator_t = typename container_t::iterator;
+  using map_t = std::unordered_map<KT, value_iterator_t>;
+  using creator_t = VT (*)();
+
+ public:
+  DNNLPrimitiveCache(size_t capacity)
+      : capacity_(capacity),
+        values_(),
+        key_to_value_(std::min(256lu, capacity)) {
+    assert(capacity > 0);
+  }
+
+  template <typename F>
+  result_value_t get_or_create(const KT& key, F&& creator) {
+    std::optional<value_iterator_t> value = get_value(key);
+    if (value.has_value()) {
+      return value.value()->second;
+    } else {
+      return add_value({key, creator()})->second;
+    }
+  }
+
+  size_t size() const { return values_.size(); }
+
+ private:
+  void dump_data() {
+    std::stringstream ss;
+    ss << "table_id: " << std::hex << reinterpret_cast<size_t>(this) << std::dec
+       << "\n";
+    ss << "container: [";
+    for (auto&& iter : values_) {
+      ss << "(" << iter.first << ", " << std::hex
+         << reinterpret_cast<size_t>(iter.second.get()) << "), " << std::dec;
+    }
+    ss << "]\n";
+
+    ss << "map: [";
+    for (auto&& iter : key_to_value_) {
+      ss << "(" << iter.first << ", " << iter.second->first << ", " << std::hex
+         << reinterpret_cast<size_t>(iter.second->second.get()) << std::dec
+         << "), ";
+    }
+    ss << "]\n";
+    std::printf("%s\n", ss.str().c_str());
+  }
+
+  value_iterator_t add_value(cache_value_t&& new_value) {
+    if (size() == capacity_) {
+      cache_value_t& last_item = values_.back();
+      key_to_value_.erase(last_item.first);
+      values_.pop_back();
+    }
+
+    auto& added_value_ = values_.emplace_front(std::move(new_value));
+    key_to_value_.emplace(added_value_.first, values_.begin());
+    return values_.begin();
+  }
+
+  std::optional<value_iterator_t> get_value(const KT& key) {
+    if (key_to_value_.size() > 0 && key == values_.begin()->first) {
+      return values_.begin();
+    }
+
+    auto value_map_iterator = key_to_value_.find(key);
+    if (value_map_iterator != key_to_value_.end()) {
+      values_.splice(values_.begin(), values_, value_map_iterator->second);
+      return value_map_iterator->second;
+    } else {
+      return {};
+    }
+  }
+
+ private:
+  const size_t capacity_;
+  container_t values_;
+  map_t key_to_value_;
+};
+
+DNNLMatMulPrimitiveHandler::DNNLMatMulPrimitiveHandler(
+    const Args& args, dnnl::memory::data_type b_type)
+    : b_n_size_(args.b_n_size),
+      b_n_stride_(args.b_n_stride),
+      b_k_size_(args.b_k_size),
+      b_k_stride_(args.b_k_stride),
+      b_type_(b_type),
+      c_type_(args.c_type),
+      runtime_memory_ptrs_(8),
+      primitive_cache_size_(args.primitive_cache_size) {
+  assert(primitive_cache_size_ > 0);
+}
+
+void DNNLMatMulPrimitiveHandler::prepack_weight(
+    void* original_b_ptr, dnnl::memory::desc b_target_mem_desc) {
+  dnnl::memory::desc original_b_md({b_k_size_, b_n_size_}, b_type_,
+                                   {b_k_stride_, b_n_stride_});
+  dnnl::memory original_weight(original_b_md, default_engine(), original_b_ptr);
+  dnnl::memory packed_weight(b_target_mem_desc, default_engine());
+  {
+    dnnl::reorder(original_weight, packed_weight)
+        .execute(default_stream(), original_weight, packed_weight);
+    default_stream().wait();
+  }
+  memory_cache_[DNNL_ARG_WEIGHTS] = packed_weight;
+  b_target_mem_desc_ = b_target_mem_desc;
+}
+
+void DNNLMatMulPrimitiveHandler::set_runtime_memory_ptr(
+    size_t index, dnnl_memory* memory_ptr) {
+  dnnl::impl::memory_storage_t* mem_storage_ptr = memory_ptr->memory_storage();
+  dnnl_memory_desc* mem_desc = const_cast<dnnl_memory_desc*>(memory_ptr->md());
+  runtime_memory_ptrs_[index] = {mem_storage_ptr, mem_desc};
+}
+
+std::pair<dnnl::impl::memory_storage_t*, dnnl_memory_desc*>
+DNNLMatMulPrimitiveHandler::get_runtime_memory_ptr(size_t index) {
+  return runtime_memory_ptrs_[index];
+}
+
+namespace std {
+template <>
+struct hash<W8A8MatMulPrimitiveHandler::ClassMatmulCacheKey> {
+  size_t operator()(
+      const W8A8MatMulPrimitiveHandler::ClassMatmulCacheKey& val) const {
+    return hash<dnnl_dim_t>()(val.b_n_size) ^ hash<dnnl_dim_t>()(val.b_k_size) ^
+           hash<int>()(static_cast<int>(val.a_qs)) ^
+           hash<int>()(static_cast<int>(val.b_qs)) ^ hash<bool>()(val.use_azp) ^
+           hash<int>()(static_cast<int>(val.c_type));
+  }
+};
+
+template <>
+struct hash<W8A8MatMulPrimitiveHandler::MSizeCacheKey> {
+  size_t operator()(
+      const W8A8MatMulPrimitiveHandler::MSizeCacheKey& val) const {
+    return hash<dnnl_dim_t>()(val.a_m_size) ^ hash<bool>()(val.use_bias) ^
+           hash<int>()(static_cast<int>(val.bias_type));
+  }
+};
+}  // namespace std
+
+bool operator==(const W8A8MatMulPrimitiveHandler::ClassMatmulCacheKey& l,
+                const W8A8MatMulPrimitiveHandler::ClassMatmulCacheKey& r) {
+  return l.b_n_size == r.b_n_size && l.b_k_size == r.b_k_size &&
+         l.a_qs == r.a_qs && l.b_qs == r.b_qs && l.use_azp == r.use_azp &&
+         l.c_type == r.c_type;
+}
+
+bool operator==(const W8A8MatMulPrimitiveHandler::MSizeCacheKey& l,
+                const W8A8MatMulPrimitiveHandler::MSizeCacheKey& r) {
+  return l.use_bias == r.use_bias && l.a_m_size == r.a_m_size &&
+         l.bias_type == r.bias_type;
+}
+
+static std::shared_ptr<W8A8MatMulPrimitiveHandler::MSizeCache>
+get_w8a8_class_primitive_cache(
+    const W8A8MatMulPrimitiveHandler::ClassMatmulCacheKey& key,
+    int64_t cache_size) {
+  static W8A8MatMulPrimitiveHandler::ClassMatmulCache cache(128);
+  assert(cache_size > 0);
+  return cache.get_or_create(key, [&]() {
+    return std::make_shared<W8A8MatMulPrimitiveHandler::MSizeCache>(cache_size);
+  });
+}
+
+W8A8MatMulPrimitiveHandler::W8A8MatMulPrimitiveHandler(const Args& args)
+    : DNNLMatMulPrimitiveHandler(
+          static_cast<const DNNLMatMulPrimitiveHandler::Args&>(args),
+          dnnl::memory::data_type::s8),
+      use_azp_(args.use_a_zero_point),
+      a_qs_(args.a_quantization_strategy),
+      b_qs_(args.b_quantization_strategy),
+      m_size_cache_(nullptr) {
+  assert(a_qs_ != QuantizationStrategy::PER_OUTPUT_CHANNEL);
+  assert(b_qs_ != QuantizationStrategy::PER_TOKEN);
+  if (a_qs_ == QuantizationStrategy::PER_TOKEN) {
+    assert(!use_azp_);
+  };
+  prepack_weight(args.b_ptr,
+                 create_primitive_desc(
+                     MSizeCacheKey{.a_m_size = DNNL_RUNTIME_DIM_VAL,
+                                   .use_bias = false,
+                                   .bias_type = dnnl::memory::data_type::undef},
+                     true)
+                     .weights_desc());
+  init_runtime_memory_cache(args);
+}
+
+void W8A8MatMulPrimitiveHandler::execute(ExecArgs& args) {
+  auto&& [a_storage, a_mem_desc] = get_runtime_memory_ptr(0);
+  auto&& [c_storage, c_mem_desc] = get_runtime_memory_ptr(1);
+  a_storage->set_data_handle((void*)args.a_ptr);
+  a_mem_desc->dims[0] = args.a_m_size;
+  c_storage->set_data_handle((void*)args.c_ptr);
+  c_mem_desc->dims[0] = args.a_m_size;
+
+  if (a_qs_ == QuantizationStrategy::PER_TENSOR) {
+    auto&& [a_scale_storage, a_scale_mem_desc] = get_runtime_memory_ptr(2);
+    a_scale_storage->set_data_handle((void*)args.a_scales_ptr);
+  }
+  if (use_azp_) {
+    auto&& [a_zero_point_storage, a_zero_point_mem_desc] =
+        get_runtime_memory_ptr(3);
+    a_zero_point_storage->set_data_handle((void*)args.a_zero_points_ptr);
+  }
+
+  if (args.use_bias) {
+    auto&& [bias_storage, bias_mem_desc] = get_runtime_memory_ptr(4);
+    bias_storage->set_data_handle((void*)args.bias_ptr);
+  }
+
+  dnnl::matmul matmul = get_matmul_cache(args);
+  matmul.execute(default_stream(), memory_cache_);
+  default_stream().wait();
+}
+
+dnnl::matmul W8A8MatMulPrimitiveHandler::get_matmul_cache(
+    const MSizeCacheKey& key) {
+  if (m_size_cache_.get() == nullptr) {
+    ClassMatmulCacheKey key = {.b_n_size = b_n_size_,
+                               .b_k_size = b_k_size_,
+                               .a_qs = a_qs_,
+                               .b_qs = b_qs_,
+                               .use_azp = use_azp_,
+                               .c_type = c_type_};
+    m_size_cache_ = get_w8a8_class_primitive_cache(key, primitive_cache_size_);
+  }
+
+  return m_size_cache_->get_or_create(key, [&]() {
+    dnnl::matmul::primitive_desc desc = this->create_primitive_desc(key, false);
+    return dnnl::matmul(desc);
+  });
+}
+
+void W8A8MatMulPrimitiveHandler::init_runtime_memory_cache(const Args& args) {
+  memory_cache_[DNNL_ARG_SRC] = dnnl::memory({{1, b_k_size_},
+                                              dnnl::memory::data_type::s8,
+                                              dnnl::memory::format_tag::ab},
+                                             default_engine(), nullptr);
+  set_runtime_memory_ptr(0, memory_cache_[DNNL_ARG_SRC].get());
+  memory_cache_[DNNL_ARG_DST] =
+      dnnl::memory({{1, b_n_size_}, c_type_, dnnl::memory::format_tag::ab},
+                   default_engine(), nullptr);
+  set_runtime_memory_ptr(1, memory_cache_[DNNL_ARG_DST].get());
+
+  // For PER_TOKEN, scales will be applied in outside epilogue
+  if (a_qs_ == QuantizationStrategy::PER_TENSOR) {
+    memory_cache_[DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC] = dnnl::memory(
+        {{1}, dnnl::memory::data_type::f32, {1}}, default_engine(), nullptr);
+    set_runtime_memory_ptr(
+        2, memory_cache_[DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC].get());
+    if (use_azp_) {
+      memory_cache_[DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_SRC] = dnnl::memory(
+          {{1}, dnnl::memory::data_type::s32, {1}}, default_engine(), nullptr);
+      set_runtime_memory_ptr(
+          3, memory_cache_[DNNL_ARG_ATTR_ZERO_POINTS | DNNL_ARG_SRC].get());
+    }
+  }
+
+  if (b_qs_ == QuantizationStrategy::PER_TENSOR) {
+    memory_cache_[DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS] =
+        dnnl::memory({{1}, dnnl::memory::data_type::f32, {1}}, default_engine(),
+                     (void*)args.b_scales_ptr);
+  } else if (b_qs_ == QuantizationStrategy::PER_OUTPUT_CHANNEL) {
+    memory_cache_[DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS] =
+        dnnl::memory({{b_n_size_}, dnnl::memory::data_type::f32, {1}},
+                     default_engine(), (void*)args.b_scales_ptr);
+  }
+
+  memory_cache_[DNNL_ARG_BIAS] =
+      dnnl::memory({{b_n_size_}, dnnl::memory::data_type::f32, {1}},
+                   default_engine(), nullptr);
+  set_runtime_memory_ptr(4, memory_cache_[DNNL_ARG_BIAS].get());
+}
+
+dnnl::matmul::primitive_desc W8A8MatMulPrimitiveHandler::create_primitive_desc(
+    const MSizeCacheKey& key, bool first_time) {
+  dnnl::memory::desc a_md({key.a_m_size, b_k_size_},
+                          dnnl::memory::data_type::s8,
+                          dnnl::memory::format_tag::ab);
+  dnnl::memory::desc b_md;
+  if (first_time) {
+    b_md =
+        dnnl::memory::desc({b_k_size_, b_n_size_}, dnnl::memory::data_type::s8,
+                           dnnl::memory::format_tag::any);
+  } else {
+    b_md = b_target_mem_desc_;
+  }
+  dnnl::memory::desc c_md({key.a_m_size, b_n_size_}, c_type_,
+                          dnnl::memory::format_tag::ab);
+
+  dnnl::primitive_attr attr;
+  // For PER_TOKEN, scales will be applied in outside epilogue
+  if (a_qs_ == QuantizationStrategy::PER_TENSOR) {
+    attr.set_scales_mask(DNNL_ARG_SRC, 0);
+    if (use_azp_) {
+      attr.set_zero_points_mask(DNNL_ARG_SRC, 0);
+    }
+  }
+
+  if (b_qs_ == QuantizationStrategy::PER_TENSOR) {
+    attr.set_scales_mask(DNNL_ARG_WEIGHTS, 0);
+  } else if (b_qs_ == QuantizationStrategy::PER_OUTPUT_CHANNEL) {
+    attr.set_scales_mask(DNNL_ARG_WEIGHTS, 2);
+  }
+
+  if (key.use_bias) {
+    // For PER_TOKEN, bias will be applied in epilogue
+    assert(a_qs_ == QuantizationStrategy::PER_TENSOR);
+    dnnl::memory::desc bias_md({1, b_n_size_}, key.bias_type, {b_n_size_, 1});
+    return dnnl::matmul::primitive_desc(default_engine(), a_md, b_md, bias_md,
+                                        c_md, attr);
+  } else {
+    return dnnl::matmul::primitive_desc(default_engine(), a_md, b_md, c_md,
+                                        attr);
+  }
+}

csrc/cpu/dnnl_helper.h

+#ifndef DNNL_HELPER_H
+#define DNNL_HELPER_H
+
+#include <optional>
+#include <cassert>
+
+#include "oneapi/dnnl/dnnl.hpp"
+
+namespace c10 {
+struct BFloat16;
+struct Half;
+}  // namespace c10
+
+namespace dnnl {
+namespace impl {
+struct memory_storage_t;
+struct matmul_pd_t;
+struct matmul_desc_t;
+}  // namespace impl
+}  // namespace dnnl
+struct dnnl_memory_desc;
+
+template <typename KT, typename VT>
+class DNNLPrimitiveCache;
+
+template <typename T>
+struct DNNLType {
+  static constexpr dnnl::memory::data_type type =
+      dnnl::memory::data_type::undef;
+};
+
+template <>
+struct DNNLType<int8_t> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::s8;
+};
+
+template <>
+struct DNNLType<int32_t> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::s32;
+};
+
+template <>
+struct DNNLType<float> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::f32;
+};
+
+template <>
+struct DNNLType<c10::BFloat16> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::bf16;
+};
+
+template <>
+struct DNNLType<c10::Half> {
+  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::f16;
+};
+
+template <typename T>
+constexpr inline dnnl::memory::data_type get_dnnl_type() {
+  return DNNLType<std::decay_t<T>>::type;
+}
+
+class DNNLMatMulPrimitiveHandler {
+ public:
+  virtual ~DNNLMatMulPrimitiveHandler() = default;
+
+ protected:
+  struct Args {
+    dnnl_dim_t b_n_size;
+    dnnl_dim_t b_n_stride;
+    dnnl_dim_t b_k_size;
+    dnnl_dim_t b_k_stride;
+    void* b_ptr;
+    dnnl::memory::data_type c_type;
+    size_t primitive_cache_size;
+  };
+
+ protected:
+  DNNLMatMulPrimitiveHandler(const Args& args, dnnl::memory::data_type b_type);
+
+  void prepack_weight(void* original_b_ptr,
+                      dnnl::memory::desc b_target_mem_desc);
+
+  void set_runtime_memory_ptr(size_t index, dnnl_memory* memory_ptr);
+
+  std::pair<dnnl::impl::memory_storage_t*, dnnl_memory_desc*>
+  get_runtime_memory_ptr(size_t index);
+
+ protected:
+  const dnnl_dim_t b_n_size_;
+  const dnnl_dim_t b_n_stride_;
+  const dnnl_dim_t b_k_size_;
+  const dnnl_dim_t b_k_stride_;
+  dnnl::memory::data_type b_type_;
+  dnnl::memory::data_type c_type_;
+  std::unordered_map<int, dnnl::memory> memory_cache_;
+  std::vector<std::pair<dnnl::impl::memory_storage_t*, dnnl_memory_desc*>>
+      runtime_memory_ptrs_;
+  dnnl::memory::desc b_target_mem_desc_;
+  int64_t primitive_cache_size_;
+};
+
+class W8A8MatMulPrimitiveHandler : public DNNLMatMulPrimitiveHandler {
+ public:
+  enum class QuantizationStrategy { PER_TOKEN, PER_TENSOR, PER_OUTPUT_CHANNEL };
+
+  struct Args : public DNNLMatMulPrimitiveHandler::Args {
+    bool use_a_zero_point;
+    QuantizationStrategy a_quantization_strategy;
+    QuantizationStrategy b_quantization_strategy;
+    float* b_scales_ptr;
+  };
+
+  struct ClassMatmulCacheKey {
+    dnnl_dim_t b_n_size;
+    dnnl_dim_t b_k_size;
+    QuantizationStrategy a_qs;
+    QuantizationStrategy b_qs;
+    bool use_azp;
+    dnnl::memory::data_type c_type;
+
+    friend bool operator==(const ClassMatmulCacheKey& l,
+                           const ClassMatmulCacheKey& r);
+  };
+
+  struct MSizeCacheKey {
+    dnnl_dim_t a_m_size;
+    bool use_bias;
+    dnnl::memory::data_type bias_type;
+
+    friend bool operator==(const MSizeCacheKey& l, const MSizeCacheKey& r);
+  };
+
+  using MSizeCache = DNNLPrimitiveCache<MSizeCacheKey, dnnl::matmul>;
+  using ClassMatmulCache =
+      DNNLPrimitiveCache<ClassMatmulCacheKey, std::shared_ptr<MSizeCache>>;
+
+  struct ExecArgs : public MSizeCacheKey {
+    const int8_t* a_ptr;
+    const float* a_scales_ptr;
+    const int32_t* a_zero_points_ptr;
+    const void* bias_ptr;
+    void* c_ptr;
+  };
+
+ public:
+  W8A8MatMulPrimitiveHandler(const Args& args);
+
+  QuantizationStrategy get_input_scale_strategy() const { return a_qs_; }
+
+  bool get_input_use_zero_point() const { return use_azp_; }
+
+  void execute(ExecArgs& args);
+
+ private:
+  dnnl::matmul::primitive_desc create_primitive_desc(const MSizeCacheKey& key,
+                                                     bool first_time);
+
+  void init_runtime_memory_cache(const Args& args);
+
+  dnnl::matmul get_matmul_cache(const MSizeCacheKey& key);
+
+ private:
+  const bool use_azp_;
+  const QuantizationStrategy a_qs_;
+  const QuantizationStrategy b_qs_;
+  std::shared_ptr<MSizeCache> m_size_cache_;
+};
+
+#endif

csrc/cpu/dnnl_helper.hpp

-#ifndef DNNL_HELPER_HPP
-#define DNNL_HELPER_HPP
-
-#include <c10/util/BFloat16.h>
-#include <c10/util/Half.h>
-
-#include "oneapi/dnnl/dnnl.hpp"
-
-namespace {
-template <typename T>
-struct DNNLType {
-  static constexpr dnnl::memory::data_type type =
-      dnnl::memory::data_type::undef;
-};
-
-template <>
-struct DNNLType<int8_t> {
-  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::s8;
-};
-
-template <>
-struct DNNLType<int32_t> {
-  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::s32;
-};
-
-template <>
-struct DNNLType<float> {
-  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::f32;
-};
-
-template <>
-struct DNNLType<c10::BFloat16> {
-  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::bf16;
-};
-
-template <>
-struct DNNLType<c10::Half> {
-  static constexpr dnnl::memory::data_type type = dnnl::memory::data_type::f16;
-};
-
-template <typename T>
-constexpr inline dnnl::memory::data_type get_dnnl_type() {
-  return DNNLType<std::decay_t<T>>::type;
-}
-};  // namespace
-
-template <bool InputNoScale>
-class DNNLPrimitiveHelper {
- public:
-  // I8 input GEMM kernel (C = a_scales * A @ (b_scales * B^T) + bias)
-  // A: [M, K], row-major
-  // B: [K, N], column-major
-  // C: [M, N], row-major
-  // bias: [N], row-major, optional
-  // a_scales: [MS]
-  // b_scales: [NS]
-  // Note: Due to the limitation of oneDNN
-  // (https://github.com/oneapi-src/oneDNN/issues/1636), the quantized bias is
-  // not supported.
-
-  template <typename OutputT, typename BiasT>
-  static void gemm_s8s8_jit(const int8_t* a, const int8_t* b, OutputT* c,
-                            const BiasT* bias, dnnl_dim_t M, dnnl_dim_t N,
-                            dnnl_dim_t K, const float* a_scales,
-                            const float* b_scales, dnnl_dim_t MS,
-                            dnnl_dim_t NS) {
-    auto&& OutputType = get_dnnl_type<OutputT>();
-    auto&& BiasType = get_dnnl_type<BiasT>();
-
-    dnnl::memory::desc a_md({M, K}, dnnl::memory::data_type::s8, {K, 1});
-    dnnl::memory::desc b_md({K, N}, dnnl::memory::data_type::s8, {1, K});
-    dnnl::memory::desc c_md({M, N}, OutputType, {N, 1});
-
-    dnnl::primitive_attr attr;
-    if constexpr (!InputNoScale) {
-      if (MS == 1) {
-        // per-tensor
-        attr.set_scales_mask(DNNL_ARG_SRC, 0);
-      } else {
-        // per-token
-        TORCH_CHECK(false, "per-token quantization is unsupported.");
-      }
-    }
-
-    if (NS == 1) {
-      // per-tensor
-      attr.set_scales_mask(DNNL_ARG_WEIGHTS, 0);
-    } else {
-      // per-channel
-      attr.set_scales_mask(DNNL_ARG_WEIGHTS, 2);
-    }
-
-    dnnl::matmul::primitive_desc matmul_pd;
-// Create memory descriptors with format_tag::any for the primitive. This
-// enables the matmul primitive to choose memory layouts for an
-// optimized primitive implementation, and these layouts may differ from the
-// ones provided by the user.
-#ifdef __aarch64__
-    auto mat_src_md = dnnl::memory::desc({M, K}, dnnl::memory::data_type::s8,
-                                         dnnl::memory::format_tag::any);
-    auto mat_weights_md = dnnl::memory::desc(
-        {K, N}, dnnl::memory::data_type::s8, dnnl::memory::format_tag::any);
-    auto mat_dst_md =
-        dnnl::memory::desc({M, N}, OutputType, dnnl::memory::format_tag::any);
-    if (bias) {
-      dnnl::memory::desc bias_md({1, N}, BiasType, {N, 1});
-      matmul_pd = dnnl::matmul::primitive_desc(default_engine(), mat_src_md,
-                                               mat_weights_md, bias_md,
-                                               mat_dst_md, attr);
-    } else {
-      matmul_pd = dnnl::matmul::primitive_desc(
-          default_engine(), mat_src_md, mat_weights_md, mat_dst_md, attr);
-    }
-#else
-    if (bias) {
-      dnnl::memory::desc bias_md({1, N}, BiasType, {N, 1});
-      matmul_pd = dnnl::matmul::primitive_desc(default_engine(), a_md, b_md,
-                                               bias_md, c_md, attr);
-    } else {
-      matmul_pd = dnnl::matmul::primitive_desc(default_engine(), a_md, b_md,
-                                               c_md, attr);
-    }
-#endif
-    dnnl::matmul matmul(matmul_pd);
-
-    auto& engine = default_engine();
-
-    dnnl::memory a_m(a_md, engine, (void*)a);
-    dnnl::memory b_m(b_md, engine, (void*)b);
-    dnnl::memory c_m(c_md, engine, (void*)c);
-    dnnl::memory a_scales_m({{MS}, dnnl::memory::data_type::f32, {1}}, engine,
-                            (void*)a_scales);
-    dnnl::memory b_scales_m({{NS}, dnnl::memory::data_type::f32, {1}}, engine,
-                            (void*)b_scales);
-
-    auto& stream = default_stream();
-
-    auto mat_src_mem = a_m;
-    auto mat_weights_mem = b_m;
-    auto mat_dst_mem = c_m;
-#ifdef __aarch64__
-    if (matmul_pd.weights_desc() != b_m.get_desc()) {
-      mat_weights_mem = dnnl::memory(matmul_pd.weights_desc(), engine);
-      dnnl::reorder(b_m, mat_weights_mem).execute(stream, b_m, mat_weights_mem);
-    }
-#endif
-    if constexpr (InputNoScale) {
-      if (bias) {
-        dnnl::memory::desc bias_md({N}, BiasType, {1});
-        dnnl::memory bias_m(bias_md, engine, (void*)bias);
-        matmul.execute(
-            stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
-                        {DNNL_ARG_BIAS, bias_m},
-                        {DNNL_ARG_DST, mat_dst_mem},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
-                    });
-      } else {
-        matmul.execute(
-            stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
-                        {DNNL_ARG_DST, mat_dst_mem},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
-                    });
-      }
-    } else {
-      if (bias) {
-        dnnl::memory::desc bias_md({N}, BiasType, {1});
-        dnnl::memory bias_m(bias_md, engine, (void*)bias);
-        matmul.execute(
-            stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
-                        {DNNL_ARG_BIAS, bias_m},
-                        {DNNL_ARG_DST, mat_dst_mem},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC, a_scales_m},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
-                    });
-      } else {
-        matmul.execute(
-            stream, {
-                        {DNNL_ARG_SRC, mat_src_mem},
-                        {DNNL_ARG_WEIGHTS, mat_weights_mem},
-                        {DNNL_ARG_DST, mat_dst_mem},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC, a_scales_m},
-                        {DNNL_ARG_ATTR_SCALES | DNNL_ARG_WEIGHTS, b_scales_m},
-                    });
-      }
-    }
-    stream.wait();
-  }
-
- private:
-  static dnnl::engine& default_engine() {
-    static dnnl::engine engine(dnnl::engine::kind::cpu, 0);
-    return engine;
-  }
-
-  static dnnl::stream& default_stream() {
-    static dnnl::stream stream(default_engine());
-    return stream;
-  }
-};
-#endif

csrc/cpu/dnnl_kernels.cpp

+#include "cpu_types.hpp"
+#include "dnnl_helper.h"
+
+namespace {
+template <typename scalar_t>
+struct KernelVecType {
+  using load_vec_type = void;
+  using cvt_vec_type = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using load_vec_type = vec_op::FP32Vec16;
+  using cvt_vec_type = vec_op::FP32Vec16;
+};
+
+#if !defined(__aarch64__) || defined(ARM_BF16_SUPPORT)
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using load_vec_type = vec_op::BF16Vec16;
+  using cvt_vec_type = vec_op::FP32Vec16;
+};
+#endif
+
+template <>
+struct KernelVecType<c10::Half> {
+#if defined(__powerpc64__) || defined(__s390x__)
+  // Power architecture-specific vector type
+  using load_vec_type = vec_op::FP32Vec16;
+#else
+  // Fallback for other architectures
+  using load_vec_type = vec_op::FP16Vec16;
+#endif
+  using cvt_vec_type = vec_op::FP32Vec16;
+};
+
+template <bool AZP, typename scalar_t>
+void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
+                                   const float* scale, const int32_t* azp,
+                                   const int64_t num_tokens,
+                                   const int64_t input_stride,
+                                   const int64_t hidden_size) {
+  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
+  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
+  constexpr int64_t vec_elem_num = load_vec_t::VEC_ELEM_NUM;
+
+  constexpr float i8_min =
+      static_cast<float>(std::numeric_limits<int8_t>::min());
+  constexpr float i8_max =
+      static_cast<float>(std::numeric_limits<int8_t>::max());
+  const cvt_vec_t inv_scale(1.0 / *scale);
+  const cvt_vec_t i8_min_vec(i8_min);
+  const cvt_vec_t i8_max_vec(i8_max);
+
+  cvt_vec_t zp_vec;
+  if constexpr (AZP) {
+    zp_vec = cvt_vec_t(static_cast<float>(*azp));
+  }
+
+#pragma omp parallel for
+  for (int64_t i = 0; i < num_tokens; ++i) {
+    int64_t j = 0;
+    const scalar_t* input_ptr = input + i * input_stride;
+    int8_t* output_ptr = output + i * hidden_size;
+    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
+      load_vec_t elems(input_ptr + j);
+      cvt_vec_t elems_fp32(elems);
+      elems_fp32 = elems_fp32 * inv_scale;
+
+      if constexpr (AZP) {
+        elems_fp32 = elems_fp32 + zp_vec;
+      }
+
+      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
+      vec_op::INT8Vec16 elems_int8(elems_fp32);
+      elems_int8.save(output_ptr + j);
+    }
+
+    load_vec_t elems(input_ptr + j);
+    cvt_vec_t elems_fp32(elems);
+    elems_fp32 = elems_fp32 * inv_scale;
+
+    if constexpr (AZP) {
+      elems_fp32 = elems_fp32 + zp_vec;
+    }
+
+    elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
+    vec_op::INT8Vec16 elems_int8(elems_fp32);
+    elems_int8.save(output_ptr + j, hidden_size - j);
+  }
+}
+
+template <bool AZP, typename scalar_t>
+void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
+                                    float* scale, int32_t* azp,
+                                    const int64_t num_tokens,
+                                    const int64_t input_stride,
+                                    const int64_t hidden_size) {
+  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
+  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
+  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
+
+  constexpr float i8_min =
+      static_cast<float>(std::numeric_limits<int8_t>::min());
+  constexpr float i8_max =
+      static_cast<float>(std::numeric_limits<int8_t>::max());
+  const cvt_vec_t i8_min_vec(i8_min);
+  const cvt_vec_t i8_max_vec(i8_max);
+
+#pragma omp parallel for
+  for (int64_t i = 0; i < num_tokens; ++i) {
+    cvt_vec_t max_value(std::numeric_limits<float>::lowest());
+    cvt_vec_t min_value(std::numeric_limits<float>::max());
+    {
+      int64_t j = 0;
+      const scalar_t* input_ptr = input + i * input_stride;
+      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
+        load_vec_t elems(input_ptr + j);
+        cvt_vec_t elems_fp32(elems);
+        if constexpr (AZP) {
+          max_value = max_value.max(elems_fp32);
+          min_value = min_value.min(elems_fp32);
+        } else {
+          max_value = max_value.max(elems_fp32.abs());
+        }
+      }
+
+      load_vec_t elems(input_ptr + j);
+      cvt_vec_t elems_fp32(elems);
+
+      if (j + vec_elem_num == hidden_size) {
+        if constexpr (AZP) {
+          max_value = max_value.max(elems_fp32);
+          min_value = min_value.min(elems_fp32);
+        } else {
+          max_value = max_value.max(elems_fp32.abs());
+        }
+      } else {
+        if constexpr (AZP) {
+          max_value = max_value.max(elems_fp32, hidden_size - j);
+          min_value = min_value.min(elems_fp32, hidden_size - j);
+        } else {
+          max_value = max_value.max(elems_fp32.abs(), hidden_size - j);
+        }
+      }
+    }
+
+    float scale_val, azp_val;
+    if constexpr (AZP) {
+      float max_scalar = max_value.reduce_max();
+      float min_scalar = min_value.reduce_min();
+      scale_val = (max_scalar - min_scalar) / 255.0f;
+      azp_val = std::nearbyint(-128.0f - min_scalar / scale_val);
+      azp[i] = azp_val;
+      scale[i] = scale_val;
+    } else {
+      scale_val = max_value.reduce_max() / 127.0f;
+      scale[i] = scale_val;
+    }
+
+    const cvt_vec_t inv_scale(1.0 / scale_val);
+    const cvt_vec_t azp_vec(azp_val);
+
+    {
+      int64_t j = 0;
+      const scalar_t* input_ptr = input + i * input_stride;
+      int8_t* output_ptr = output + i * hidden_size;
+      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
+        load_vec_t elems(input_ptr + j);
+        cvt_vec_t elems_fp32(elems);
+        elems_fp32 = (elems_fp32 * inv_scale);
+
+        if constexpr (AZP) {
+          elems_fp32 = elems_fp32 + azp_vec;
+        }
+        elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
+        vec_op::INT8Vec16 elems_int8(elems_fp32);
+        elems_int8.save(output_ptr + j);
+      }
+
+      load_vec_t elems(input_ptr + j);
+      cvt_vec_t elems_fp32(elems);
+      elems_fp32 = (elems_fp32 * inv_scale);
+
+      if constexpr (AZP) {
+        elems_fp32 = elems_fp32 + azp_vec;
+      }
+      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
+      vec_op::INT8Vec16 elems_int8(elems_fp32);
+      elems_int8.save(output_ptr + j, hidden_size - j);
+    }
+  }
+}
+
+template <bool AZP, bool Bias, typename scalar_t>
+void dynamic_quant_epilogue(const float* input, scalar_t* output,
+                            const float* a_scale, const int32_t* azp,
+                            const float* azp_adj, const scalar_t* bias,
+                            const int64_t num_tokens,
+                            const int64_t hidden_size) {
+  CPU_KERNEL_GUARD_IN(dynamic_quant_epilogue)
+  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
+  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
+  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
+
+  const int64_t thread_num = omp_get_max_threads();
+  if (num_tokens > thread_num) {
+#pragma omp parallel for
+    for (int64_t i = 0; i < num_tokens; ++i) {
+      const float* input_ptr = input + i * hidden_size;
+      scalar_t* output_ptr = output + i * hidden_size;
+      int64_t j = 0;
+      cvt_vec_t token_scale_vec(a_scale[i]);
+      cvt_vec_t token_zp_scale_vec;
+      if constexpr (AZP) {
+        float zp_scale_val = a_scale[i] * static_cast<float>(azp[i]);
+        token_zp_scale_vec = cvt_vec_t(zp_scale_val);
+      }
+      for (; j < hidden_size - vec_elem_num; ++j) {
+        cvt_vec_t elems_fp32(input_ptr + j);
+        elems_fp32 = elems_fp32 * token_scale_vec;
+        if constexpr (AZP) {
+          cvt_vec_t azp_adj_fp32(azp_adj + j);
+          elems_fp32 = elems_fp32 - azp_adj_fp32 * token_zp_scale_vec;
+        }
+        if constexpr (Bias) {
+          load_vec_t bias_vec(bias + j);
+          cvt_vec_t bias_vec_fp32(bias_vec);
+          elems_fp32 = elems_fp32 + bias_vec_fp32;
+        }
+        load_vec_t elems_out(elems_fp32);
+        elems_out.save(output_ptr + j);
+      }
+      cvt_vec_t elems_fp32(input_ptr + j);
+      elems_fp32 = elems_fp32 * token_scale_vec;
+      if constexpr (AZP) {
+        cvt_vec_t azp_adj_fp32(azp_adj + j);
+        elems_fp32 = elems_fp32 - azp_adj_fp32 * token_zp_scale_vec;
+      }
+      if constexpr (Bias) {
+        load_vec_t bias_vec(bias + j);
+        cvt_vec_t bias_vec_fp32(bias_vec);
+        elems_fp32 = elems_fp32 + bias_vec_fp32;
+      }
+      load_vec_t elems_out(elems_fp32);
+      elems_out.save(output_ptr + j, hidden_size - j);
+    }
+  } else {
+    const int64_t vec_iteration =
+        (hidden_size + vec_elem_num - 1) / vec_elem_num;
+    const int64_t vec_iteration_per_thread =
+        (vec_iteration + thread_num - 1) / thread_num;
+    const int64_t elem_num_per_thread = vec_iteration_per_thread * vec_elem_num;
+#pragma omp parallel for schedule(static, 1)
+    for (int64_t i = 0; i < thread_num; ++i) {
+      const int64_t start = elem_num_per_thread * i;
+      const int64_t end = std::min(hidden_size, elem_num_per_thread + start);
+      for (int64_t j = 0; j < num_tokens; ++j) {
+        cvt_vec_t token_scale_vec(a_scale[j]);
+        cvt_vec_t token_zp_scale_vec;
+        if constexpr (AZP) {
+          float zp_scale_val = a_scale[j] * static_cast<float>(azp[j]);
+          token_zp_scale_vec = cvt_vec_t(zp_scale_val);
+        }
+        int64_t k = start;
+        const float* input_ptr = input + j * hidden_size;
+        scalar_t* output_ptr = output + j * hidden_size;
+        for (; k < end - vec_elem_num; k += vec_elem_num) {
+          cvt_vec_t elems_fp32(input_ptr + k);
+          elems_fp32 = elems_fp32 * token_scale_vec;
+          if constexpr (AZP) {
+            cvt_vec_t azp_adj_fp32(azp_adj + k);
+            elems_fp32 = elems_fp32 - azp_adj_fp32 * token_zp_scale_vec;
+          }
+          if constexpr (Bias) {
+            load_vec_t bias_vec(bias + k);
+            cvt_vec_t bias_vec_fp32(bias_vec);
+            elems_fp32 = elems_fp32 + bias_vec_fp32;
+          }
+          load_vec_t elems_out(elems_fp32);
+          elems_out.save(output_ptr + k);
+        }
+        if (k < end) {
+          cvt_vec_t elems_fp32(input_ptr + k);
+          elems_fp32 = elems_fp32 * token_scale_vec;
+          if constexpr (AZP) {
+            cvt_vec_t azp_adj_fp32(azp_adj + k);
+            elems_fp32 = elems_fp32 - azp_adj_fp32 * token_zp_scale_vec;
+          }
+          if constexpr (Bias) {
+            load_vec_t bias_vec(bias + k);
+            cvt_vec_t bias_vec_fp32(bias_vec);
+            elems_fp32 = elems_fp32 + bias_vec_fp32;
+          }
+          load_vec_t elems_out(elems_fp32);
+          elems_out.save(output_ptr + k, end - k);
+        }
+      }
+    }
+  }
+}
+}  // namespace
+
+int64_t create_onednn_scaled_mm_handler(
+    const torch::Tensor& b,         // [IC, OC], column-major
+    const torch::Tensor& b_scales,  // [1] or [OC]
+    at::ScalarType output_type, bool dynamic_act_quant, bool use_azp,
+    int64_t primitive_cache_size) {
+  TORCH_CHECK(b.dim() == 2);
+  TORCH_CHECK(b.stride(0) == 1);  // Column-major
+  TORCH_CHECK(b_scales.is_contiguous());
+
+  W8A8MatMulPrimitiveHandler::Args args;
+  args.primitive_cache_size = primitive_cache_size;
+
+  if (b_scales.numel() == 1) {
+    args.b_quantization_strategy =
+        W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TENSOR;
+  } else {
+    TORCH_CHECK_EQ(b_scales.numel(), b.size(1));
+    args.b_quantization_strategy =
+        W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_OUTPUT_CHANNEL;
+  }
+  args.b_scales_ptr = b_scales.data_ptr<float>();
+  args.b_k_size = b.size(0);
+  args.b_k_stride = b.stride(0);
+  args.b_n_size = b.size(1);
+  args.b_n_stride = b.stride(1);
+  args.b_ptr = b.data_ptr<int8_t>();
+
+  if (dynamic_act_quant) {
+    // dynamic per-token, bias, A scales and A zps will be applied in outside.
+    args.a_quantization_strategy =
+        W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TOKEN;
+    args.use_a_zero_point = false;
+  } else {
+    // static per-tensor
+    args.a_quantization_strategy =
+        W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TENSOR;
+    args.use_a_zero_point = use_azp;
+  }
+
+  VLLM_DISPATCH_FLOATING_TYPES(output_type, "create_onednn_scaled_mm_handler",
+                               [&] {
+                                 if (dynamic_act_quant) {
+                                   args.c_type = get_dnnl_type<float>();
+                                 } else {
+                                   args.c_type = get_dnnl_type<scalar_t>();
+                                 }
+                               });
+
+  return reinterpret_cast<int64_t>(new W8A8MatMulPrimitiveHandler(args));
+}
+
+void onednn_scaled_mm(
+    torch::Tensor& c,                             // [M, OC], row-major
+    const torch::Tensor& a,                       // [M, IC], row-major
+    const torch::Tensor& a_scales,                // [M] or [1]
+    const std::optional<torch::Tensor>& azp,      // [M] or [1]
+    const std::optional<torch::Tensor>& azp_adj,  // [M] or [1]
+    const std::optional<torch::Tensor>& bias,     // [N]
+    int64_t handler) {
+  CPU_KERNEL_GUARD_IN(onednn_scaled_mm)
+  TORCH_CHECK(a.dim() == 2);
+  TORCH_CHECK(a.is_contiguous());
+  TORCH_CHECK(c.is_contiguous());
+  W8A8MatMulPrimitiveHandler* ptr =
+      reinterpret_cast<W8A8MatMulPrimitiveHandler*>(handler);
+  const int32_t* azp_ptr = nullptr;
+  if (azp.has_value()) {
+    azp_ptr = azp->data_ptr<int32_t>();
+  }
+  if (ptr->get_input_scale_strategy() ==
+      W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TENSOR) {
+    TORCH_CHECK_EQ(a_scales.numel(), 1);
+  }
+
+  W8A8MatMulPrimitiveHandler::ExecArgs exec_args;
+  exec_args.a_ptr = a.data_ptr<int8_t>();
+  exec_args.a_m_size = a.size(0);
+  exec_args.bias_ptr = nullptr;
+  exec_args.use_bias = false;
+  exec_args.a_scales_ptr = nullptr;
+  exec_args.a_zero_points_ptr = nullptr;
+
+  VLLM_DISPATCH_FLOATING_TYPES(c.scalar_type(), "onednn_scaled_mm", [&] {
+    if (ptr->get_input_scale_strategy() ==
+        W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TENSOR) {
+      if (bias.has_value()) {
+        exec_args.bias_ptr = bias->data_ptr<scalar_t>();
+        exec_args.bias_type = get_dnnl_type<scalar_t>();
+        exec_args.use_bias = true;
+      }
+      exec_args.a_scales_ptr = a_scales.data_ptr<float>();
+      exec_args.a_zero_points_ptr = azp_ptr;
+      exec_args.c_ptr = c.data_ptr<scalar_t>();
+      ptr->execute(exec_args);
+    } else if (ptr->get_input_scale_strategy() ==
+               W8A8MatMulPrimitiveHandler::QuantizationStrategy::PER_TOKEN) {
+      torch::Tensor tmp_fp32_out =
+          torch::empty_like(c, ::at::ScalarType::Float);
+      exec_args.c_ptr = tmp_fp32_out.data_ptr<float>();
+      ptr->execute(exec_args);
+      if (bias.has_value()) {
+        if (azp.has_value()) {
+          dynamic_quant_epilogue<true, true>(
+              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
+              a_scales.data_ptr<float>(), azp_ptr, azp_adj->data_ptr<float>(),
+              bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
+        } else {
+          dynamic_quant_epilogue<false, true>(
+              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
+              a_scales.data_ptr<float>(), azp_ptr, nullptr,
+              bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
+        }
+      } else {
+        if (azp.has_value()) {
+          dynamic_quant_epilogue<true, false>(
+              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
+              a_scales.data_ptr<float>(), azp_ptr, azp_adj->data_ptr<float>(),
+              (scalar_t*)nullptr, c.size(0), c.size(1));
+        } else {
+          dynamic_quant_epilogue<false, false>(
+              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
+              a_scales.data_ptr<float>(), azp_ptr, nullptr, (scalar_t*)nullptr,
+              c.size(0), c.size(1));
+        }
+      }
+    } else {
+      TORCH_CHECK(false, "invalid act quant type.");
+    }
+  });
+}
+
+// static-per-tensor quantization.
+void static_scaled_int8_quant(
+    torch::Tensor& out,          // [batch, hidden_size]
+    const torch::Tensor& input,  // [batch, hidden_size]
+    const torch::Tensor& scale, std::optional<torch::Tensor> const& azp) {
+  CPU_KERNEL_GUARD_IN(static_scaled_int8_quant)
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK_EQ(input.dim(), 2);
+  TORCH_CHECK_EQ(input.stride(1), 1);
+  TORCH_CHECK(scale.numel() == 1);
+  TORCH_CHECK(!azp.has_value() || azp->numel() == 1);
+
+  const int64_t stride = input.stride(0);
+  const int64_t hidden_size = input.size(1);
+  const int64_t num_tokens = input.size(0);
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "static_scaled_int8_quant_impl", [&] {
+        if (azp.has_value()) {
+          static_scaled_int8_quant_impl<true>(
+              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+              scale.data_ptr<float>(), azp->data_ptr<int32_t>(), num_tokens,
+              stride, hidden_size);
+        } else {
+          static_scaled_int8_quant_impl<false>(input.data_ptr<scalar_t>(),
+                                               out.data_ptr<int8_t>(),
+                                               scale.data_ptr<float>(), nullptr,
+                                               num_tokens, stride, hidden_size);
+        }
+      });
+}
+
+// dynamic-per-token quantization.
+void dynamic_scaled_int8_quant(
+    torch::Tensor& out,          // [batch, hidden_size]
+    const torch::Tensor& input,  // [batch, hidden_size]
+    torch::Tensor& scale,        // [batch, 1]
+    std::optional<torch::Tensor> const& azp) {
+  CPU_KERNEL_GUARD_IN(dynamic_scaled_int8_quant)
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK_EQ(input.dim(), 2);
+  TORCH_CHECK_EQ(input.stride(1), 1);
+
+  const int64_t hidden_size = input.size(1);
+  const int64_t num_tokens = input.size(0);
+  const int64_t stride = input.stride(0);
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "dynamic_scaled_int8_quant_impl", [&] {
+        if (azp.has_value()) {
+          dynamic_scaled_int8_quant_impl<true>(
+              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+              scale.data_ptr<float>(), azp->data_ptr<int32_t>(), num_tokens,
+              stride, hidden_size);
+        } else {
+          dynamic_scaled_int8_quant_impl<false>(
+              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
+              scale.data_ptr<float>(), nullptr, num_tokens, stride,
+              hidden_size);
+        }
+      });
+}

csrc/cpu/quant.cpp

-#include "cpu_types.hpp"
-#include "dnnl_helper.hpp"
-
-namespace {
-template <typename scalar_t>
-struct KernelVecType {
-  using load_vec_type = void;
-  using azp_adj_load_vec_type = void;
-  using cvt_vec_type = void;
-};
-
-template <>
-struct KernelVecType<float> {
-  using load_vec_type = vec_op::FP32Vec16;
-  using azp_adj_load_vec_type = vec_op::INT32Vec16;
-  using cvt_vec_type = vec_op::FP32Vec16;
-};
-
-#if !defined(__aarch64__) || defined(ARM_BF16_SUPPORT)
-template <>
-struct KernelVecType<c10::BFloat16> {
-  using load_vec_type = vec_op::BF16Vec16;
-  using azp_adj_load_vec_type = vec_op::INT32Vec16;
-  using cvt_vec_type = vec_op::FP32Vec16;
-};
-#endif
-
-template <>
-struct KernelVecType<c10::Half> {
-#if defined(__powerpc64__) || defined(__s390x__)
-  // Power architecture-specific vector type
-  using load_vec_type = vec_op::FP32Vec16;
-#else
-  // Fallback for other architectures
-  using load_vec_type = vec_op::FP16Vec16;
-#endif
-  using azp_adj_load_vec_type = vec_op::INT32Vec16;
-  using cvt_vec_type = vec_op::FP32Vec16;
-};
-
-#if defined(__AVX512F__) || defined(__aarch64__)
-template <bool AZP, typename scalar_t>
-void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                   const float* scale, const int32_t* azp,
-                                   const int num_tokens,
-                                   const int hidden_size) {
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  constexpr float i8_min =
-      static_cast<float>(std::numeric_limits<int8_t>::min());
-  constexpr float i8_max =
-      static_cast<float>(std::numeric_limits<int8_t>::max());
-  const cvt_vec_t inv_scale(1.0 / *scale);
-  const cvt_vec_t i8_min_vec(i8_min);
-  const cvt_vec_t i8_max_vec(i8_max);
-
-  cvt_vec_t zp_vec;
-  if constexpr (AZP) {
-    zp_vec = cvt_vec_t(static_cast<float>(*azp));
-  }
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    int j = 0;
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-      elems_fp32 = elems_fp32 * inv_scale;
-
-      if constexpr (AZP) {
-        elems_fp32 = elems_fp32 + zp_vec;
-      }
-
-      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-      vec_op::INT8Vec16 elems_int8(elems_fp32);
-      elems_int8.save(output + i * hidden_size + j);
-    }
-
-    load_vec_t elems(input + i * hidden_size + j);
-    cvt_vec_t elems_fp32(elems);
-    elems_fp32 = elems_fp32 * inv_scale;
-
-    if constexpr (AZP) {
-      elems_fp32 = elems_fp32 + zp_vec;
-    }
-
-    elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-    vec_op::INT8Vec16 elems_int8(elems_fp32);
-    elems_int8.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-
-template <bool AZP, typename scalar_t>
-void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                    float* scale, int32_t* azp,
-                                    const int num_tokens,
-                                    const int hidden_size) {
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  constexpr float i8_min =
-      static_cast<float>(std::numeric_limits<int8_t>::min());
-  constexpr float i8_max =
-      static_cast<float>(std::numeric_limits<int8_t>::max());
-  const cvt_vec_t i8_min_vec(i8_min);
-  const cvt_vec_t i8_max_vec(i8_max);
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    cvt_vec_t max_value(std::numeric_limits<float>::lowest());
-    cvt_vec_t min_value(std::numeric_limits<float>::max());
-    {
-      int j = 0;
-      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-        load_vec_t elems(input + i * hidden_size + j);
-        cvt_vec_t elems_fp32(elems);
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32);
-          min_value = min_value.min(elems_fp32);
-        } else {
-          max_value = max_value.max(elems_fp32.abs());
-        }
-      }
-
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-
-      if (j + vec_elem_num == hidden_size) {
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32);
-          min_value = min_value.min(elems_fp32);
-        } else {
-          max_value = max_value.max(elems_fp32.abs());
-        }
-      } else {
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32, hidden_size - j);
-          min_value = min_value.min(elems_fp32, hidden_size - j);
-        } else {
-          max_value = max_value.max(elems_fp32.abs(), hidden_size - j);
-        }
-      }
-    }
-
-    float scale_val, azp_val;
-    if constexpr (AZP) {
-      float max_scalar = max_value.reduce_max();
-      float min_scalar = min_value.reduce_min();
-      scale_val = (max_scalar - min_scalar) / 255.0f;
-      azp_val = std::nearbyint(-128.0f - min_scalar / scale_val);
-      azp[i] = static_cast<int32_t>(azp_val);
-      scale[i] = scale_val;
-    } else {
-      scale_val = max_value.reduce_max() / 127.0f;
-      scale[i] = scale_val;
-    }
-
-    const cvt_vec_t inv_scale(1.0 / scale_val);
-    const cvt_vec_t azp_vec(azp_val);
-
-    {
-      int j = 0;
-      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-        load_vec_t elems(input + i * hidden_size + j);
-        cvt_vec_t elems_fp32(elems);
-        elems_fp32 = (elems_fp32 * inv_scale);
-
-        if constexpr (AZP) {
-          elems_fp32 = elems_fp32 + azp_vec;
-        }
-        elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-        vec_op::INT8Vec16 elems_int8(elems_fp32);
-        elems_int8.save(output + i * hidden_size + j);
-      }
-
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-      elems_fp32 = (elems_fp32 * inv_scale);
-
-      if constexpr (AZP) {
-        elems_fp32 = elems_fp32 + azp_vec;
-      }
-      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-      vec_op::INT8Vec16 elems_int8(elems_fp32);
-      elems_int8.save(output + i * hidden_size + j, hidden_size - j);
-    }
-  }
-}
-
-template <bool PerChannel, typename scalar_t>
-void static_quant_epilogue(const float* input, scalar_t* output,
-                           const float a_scale, const float* b_scale,
-                           const int32_t* azp_with_adj, const int num_tokens,
-                           const int hidden_size) {
-  CPU_KERNEL_GUARD_IN(dynamic_output_scale_impl)
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using azp_adj_load_vec_t =
-      typename KernelVecType<scalar_t>::azp_adj_load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    cvt_vec_t a_scale_vec(a_scale);
-    cvt_vec_t b_scale_vec(*b_scale);
-    cvt_vec_t scale_vec = a_scale_vec * b_scale_vec;
-
-    int j = 0;
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      cvt_vec_t elems_fp32(input + i * hidden_size + j);
-      azp_adj_load_vec_t azp_adj_vec(azp_with_adj + j);
-      cvt_vec_t azp_adj_fp32(azp_adj_vec);
-
-      if constexpr (PerChannel) {
-        b_scale_vec = cvt_vec_t(b_scale + j);
-        scale_vec = b_scale_vec * a_scale_vec;
-      }
-
-      elems_fp32 = elems_fp32 - scale_vec * azp_adj_fp32;
-
-      load_vec_t elems_out(elems_fp32);
-      elems_out.save(output + i * hidden_size + j);
-    }
-
-    cvt_vec_t elems_fp32(input + i * hidden_size + j);
-    azp_adj_load_vec_t azp_adj_vec(azp_with_adj + j);
-    cvt_vec_t azp_adj_fp32(azp_adj_vec);
-
-    if constexpr (PerChannel) {
-      b_scale_vec = cvt_vec_t(b_scale + j);
-      scale_vec = b_scale_vec * a_scale_vec;
-    }
-
-    elems_fp32 = elems_fp32 - scale_vec * azp_adj_fp32;
-
-    load_vec_t elems_out(elems_fp32);
-    elems_out.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-
-template <bool AZP, bool PerChannel, bool Bias, typename scalar_t>
-void dynamic_quant_epilogue(const float* input, scalar_t* output,
-                            const float* a_scale, const float* b_scale,
-                            const int32_t* azp, const int32_t* azp_adj,
-                            const scalar_t* bias, const int num_tokens,
-                            const int hidden_size) {
-  CPU_KERNEL_GUARD_IN(dynamic_quant_epilogue)
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using azp_adj_load_vec_t =
-      typename KernelVecType<scalar_t>::azp_adj_load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    int j = 0;
-    cvt_vec_t token_scale_vec(a_scale[i]);
-    cvt_vec_t token_zp_scale_vec;
-    if constexpr (AZP) {
-      float zp_scale_val = a_scale[i] * static_cast<float>(azp[i]);
-      if constexpr (!PerChannel) {
-        zp_scale_val *= *b_scale;
-      }
-      token_zp_scale_vec = cvt_vec_t(zp_scale_val);
-    }
-
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      cvt_vec_t elems_fp32(input + i * hidden_size + j);
-      elems_fp32 = elems_fp32 * token_scale_vec;
-
-      if constexpr (AZP) {
-        azp_adj_load_vec_t azp_adj_vec(azp_adj + j);
-        cvt_vec_t azp_adj_fp32(azp_adj_vec);
-        azp_adj_fp32 = azp_adj_fp32 * token_zp_scale_vec;
-
-        if constexpr (PerChannel) {
-          cvt_vec_t b_scale_vec(b_scale + j);
-          azp_adj_fp32 = azp_adj_fp32 * b_scale_vec;
-        }
-
-        elems_fp32 = elems_fp32 - azp_adj_fp32;
-      }
-
-      if constexpr (Bias) {
-        load_vec_t bias_vec(bias + j);
-        cvt_vec_t bias_vec_fp32(bias_vec);
-        elems_fp32 = elems_fp32 + bias_vec_fp32;
-      }
-
-      load_vec_t elems_out(elems_fp32);
-      elems_out.save(output + i * hidden_size + j);
-    }
-
-    cvt_vec_t elems_fp32(input + i * hidden_size + j);
-    elems_fp32 = elems_fp32 * token_scale_vec;
-
-    if constexpr (AZP) {
-      azp_adj_load_vec_t azp_adj_vec(azp_adj + j);
-      cvt_vec_t azp_adj_fp32(azp_adj_vec);
-      azp_adj_fp32 = azp_adj_fp32 * token_zp_scale_vec;
-
-      if constexpr (PerChannel) {
-        cvt_vec_t b_scale_vec(b_scale + j);
-        azp_adj_fp32 = azp_adj_fp32 * b_scale_vec;
-      }
-
-      elems_fp32 = elems_fp32 - azp_adj_fp32;
-    }
-
-    if constexpr (Bias) {
-      load_vec_t bias_vec(bias + j);
-      cvt_vec_t bias_vec_fp32(bias_vec);
-      elems_fp32 = elems_fp32 + bias_vec_fp32;
-    }
-
-    load_vec_t elems_out(elems_fp32);
-    elems_out.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-#elif defined(__powerpc64__)
-template <bool AZP, typename scalar_t>
-void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                   const float* scale, const int32_t* azp,
-                                   const int num_tokens,
-                                   const int hidden_size) {
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  constexpr float i8_min =
-      static_cast<float>(std::numeric_limits<int8_t>::min());
-  constexpr float i8_max =
-      static_cast<float>(std::numeric_limits<int8_t>::max());
-
-  const cvt_vec_t inv_scale(1.0 / *scale);
-  const cvt_vec_t i8_min_vec(i8_min);
-  const cvt_vec_t i8_max_vec(i8_max);
-
-  cvt_vec_t zp_vec;
-  if constexpr (AZP) {
-    zp_vec = cvt_vec_t(static_cast<float>(*azp));
-  }
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    int j = 0;
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-      elems_fp32 = elems_fp32 * inv_scale;
-      if constexpr (AZP) {
-        elems_fp32 = elems_fp32 + zp_vec;
-      }
-      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-      vec_op::INT8Vec16 elems_int8(elems_fp32);
-      elems_int8.save(output + i * hidden_size + j);
-    }
-    load_vec_t elems(input + i * hidden_size + j);
-    cvt_vec_t elems_fp32(elems);
-    elems_fp32 = elems_fp32 * inv_scale;
-
-    if constexpr (AZP) {
-      elems_fp32 = elems_fp32 + zp_vec;
-    }
-
-    elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-    vec_op::INT8Vec16 elems_int8(elems_fp32);
-    elems_int8.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-template <bool AZP, typename scalar_t>
-void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                    float* scale, int32_t* azp,
-                                    const int num_tokens,
-                                    const int hidden_size) {
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  constexpr float i8_min =
-      static_cast<float>(std::numeric_limits<int8_t>::min());
-  constexpr float i8_max =
-      static_cast<float>(std::numeric_limits<int8_t>::max());
-  const cvt_vec_t i8_min_vec(i8_min);
-  const cvt_vec_t i8_max_vec(i8_max);
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    cvt_vec_t max_value(std::numeric_limits<float>::lowest());
-    cvt_vec_t min_value(std::numeric_limits<float>::max());
-    {
-      int j = 0;
-      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-        load_vec_t elems(input + i * hidden_size + j);
-        cvt_vec_t elems_fp32(elems);
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32);
-          min_value = min_value.min(elems_fp32);
-        } else {
-          max_value = max_value.max(elems_fp32.abs());
-        }
-      }
-
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-
-      if (j + vec_elem_num == hidden_size) {
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32);
-          min_value = min_value.min(elems_fp32);
-        } else {
-          max_value = max_value.max(elems_fp32.abs());
-        }
-      } else {
-        if constexpr (AZP) {
-          max_value = max_value.max(elems_fp32, hidden_size - j);
-          min_value = min_value.min(elems_fp32, hidden_size - j);
-        } else {
-          max_value = max_value.max(elems_fp32.abs(), hidden_size - j);
-        }
-      }
-    }
-
-    float scale_val, azp_val;
-    if constexpr (AZP) {
-      float max_scalar = max_value.reduce_max();
-      float min_scalar = min_value.reduce_min();
-      scale_val = (max_scalar - min_scalar) / 255.0f;
-      azp_val = std::nearbyint(-128.0f - min_scalar / scale_val);
-      azp[i] = static_cast<int32_t>(azp_val);
-      scale[i] = scale_val;
-    } else {
-      scale_val = max_value.reduce_max() / 127.0f;
-      scale[i] = scale_val;
-    }
-
-    const cvt_vec_t inv_scale(1.0 / scale_val);
-    const cvt_vec_t azp_vec(azp_val);
-
-    {
-      int j = 0;
-      for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-        load_vec_t elems(input + i * hidden_size + j);
-        cvt_vec_t elems_fp32(elems);
-        elems_fp32 = (elems_fp32 * inv_scale);
-
-        if constexpr (AZP) {
-          elems_fp32 = elems_fp32 + azp_vec;
-        }
-        elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-        vec_op::INT8Vec16 elems_int8(elems_fp32);
-        elems_int8.save(output + i * hidden_size + j);
-      }
-
-      load_vec_t elems(input + i * hidden_size + j);
-      cvt_vec_t elems_fp32(elems);
-      elems_fp32 = (elems_fp32 * inv_scale);
-
-      if constexpr (AZP) {
-        elems_fp32 = elems_fp32 + azp_vec;
-      }
-      elems_fp32 = elems_fp32.clamp(i8_min_vec, i8_max_vec);
-      vec_op::INT8Vec16 elems_int8(elems_fp32);
-      elems_int8.save(output + i * hidden_size + j, hidden_size - j);
-    }
-  }
-}
-template <bool PerChannel, typename scalar_t>
-void static_quant_epilogue(const float* input, scalar_t* output,
-                           const float a_scale, const float* b_scale,
-                           const int32_t* azp_with_adj, const int num_tokens,
-                           const int hidden_size) {
-  CPU_KERNEL_GUARD_IN(dynamic_output_scale_impl)
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using azp_adj_load_vec_t =
-      typename KernelVecType<scalar_t>::azp_adj_load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    cvt_vec_t a_scale_vec(a_scale);
-    cvt_vec_t b_scale_vec(*b_scale);
-    cvt_vec_t scale_vec = a_scale_vec * b_scale_vec;
-
-    int j = 0;
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      cvt_vec_t elems_fp32(input + i * hidden_size + j);
-      azp_adj_load_vec_t azp_adj_vec(azp_with_adj + j);
-      cvt_vec_t azp_adj_fp32(azp_adj_vec);
-
-      if constexpr (PerChannel) {
-        b_scale_vec = cvt_vec_t(b_scale + j);
-        scale_vec = b_scale_vec * a_scale_vec;
-      }
-      elems_fp32 = elems_fp32 - scale_vec * azp_adj_fp32;
-      load_vec_t elems_out(elems_fp32);
-      elems_out.save(output + i * hidden_size + j);
-    }
-
-    cvt_vec_t elems_fp32(input + i * hidden_size + j);
-    azp_adj_load_vec_t azp_adj_vec(azp_with_adj + j);
-    cvt_vec_t azp_adj_fp32(azp_adj_vec);
-
-    if constexpr (PerChannel) {
-      b_scale_vec = cvt_vec_t(b_scale + j);
-      scale_vec = b_scale_vec * a_scale_vec;
-    }
-
-    elems_fp32 = elems_fp32 - scale_vec * azp_adj_fp32;
-
-    load_vec_t elems_out(elems_fp32);
-    elems_out.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-template <bool AZP, bool PerChannel, bool Bias, typename scalar_t>
-void dynamic_quant_epilogue(const float* input, scalar_t* output,
-                            const float* a_scale, const float* b_scale,
-                            const int32_t* azp, const int32_t* azp_adj,
-                            const scalar_t* bias, const int num_tokens,
-                            const int hidden_size) {
-  CPU_KERNEL_GUARD_IN(dynamic_quant_epilogue)
-  using load_vec_t = typename KernelVecType<scalar_t>::load_vec_type;
-  using azp_adj_load_vec_t =
-      typename KernelVecType<scalar_t>::azp_adj_load_vec_type;
-  using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
-  constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
-
-  #pragma omp parallel for
-  for (int i = 0; i < num_tokens; ++i) {
-    int j = 0;
-    cvt_vec_t token_scale_vec(a_scale[i]);
-    cvt_vec_t token_zp_scale_vec;
-    if constexpr (AZP) {
-      float zp_scale_val = a_scale[i] * static_cast<float>(azp[i]);
-      if constexpr (!PerChannel) {
-        zp_scale_val *= *b_scale;
-      }
-      token_zp_scale_vec = cvt_vec_t(zp_scale_val);
-    }
-
-    for (; j < hidden_size - vec_elem_num; j += vec_elem_num) {
-      cvt_vec_t elems_fp32(input + i * hidden_size + j);
-      elems_fp32 = elems_fp32 * token_scale_vec;
-
-      if constexpr (AZP) {
-        azp_adj_load_vec_t azp_adj_vec(azp_adj + j);
-        cvt_vec_t azp_adj_fp32(azp_adj_vec);
-        azp_adj_fp32 = azp_adj_fp32 * token_zp_scale_vec;
-
-        if constexpr (PerChannel) {
-          cvt_vec_t b_scale_vec(b_scale + j);
-          azp_adj_fp32 = azp_adj_fp32 * b_scale_vec;
-        }
-
-        elems_fp32 = elems_fp32 - azp_adj_fp32;
-      }
-
-      if constexpr (Bias) {
-        load_vec_t bias_vec(bias + j);
-        cvt_vec_t bias_vec_fp32(bias_vec);
-        elems_fp32 = elems_fp32 + bias_vec_fp32;
-      }
-
-      load_vec_t elems_out(elems_fp32);
-      elems_out.save(output + i * hidden_size + j);
-    }
-
-    cvt_vec_t elems_fp32(input + i * hidden_size + j);
-    elems_fp32 = elems_fp32 * token_scale_vec;
-
-    if constexpr (AZP) {
-      azp_adj_load_vec_t azp_adj_vec(azp_adj + j);
-      cvt_vec_t azp_adj_fp32(azp_adj_vec);
-      azp_adj_fp32 = azp_adj_fp32 * token_zp_scale_vec;
-
-      if constexpr (PerChannel) {
-        cvt_vec_t b_scale_vec(b_scale + j);
-        azp_adj_fp32 = azp_adj_fp32 * b_scale_vec;
-      }
-
-      elems_fp32 = elems_fp32 - azp_adj_fp32;
-    }
-
-    if constexpr (Bias) {
-      load_vec_t bias_vec(bias + j);
-      cvt_vec_t bias_vec_fp32(bias_vec);
-      elems_fp32 = elems_fp32 + bias_vec_fp32;
-    }
-
-    load_vec_t elems_out(elems_fp32);
-    elems_out.save(output + i * hidden_size + j, hidden_size - j);
-  }
-}
-#else
-template <typename scalar_t>
-void static_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                   const float* scale, const int32_t* azp,
-                                   const int num_tokens,
-                                   const int hidden_size) {
-  TORCH_CHECK(false,
-              "static_scaled_int8_quant_impl requires AVX512/powerpc64/AArch64 "
-              "support.")
-}
-
-template <typename scalar_t>
-void dynamic_scaled_int8_quant_impl(const scalar_t* input, int8_t* output,
-                                    float* scale, int32_t* azp,
-                                    const int num_tokens,
-                                    const int hidden_size) {
-  TORCH_CHECK(false,
-              "dynamic_scaled_int8_quant_impl requires "
-              "AVX512/powerpc64/AArch64 support.")
-}
-
-template <bool PerChannel, typename scalar_t>
-void static_quant_epilogue(const float* input, scalar_t* output,
-                           const float a_scale, const float* b_scale,
-                           const int32_t* azp_with_adj, const int num_tokens,
-                           const int hidden_size) {
-  TORCH_CHECK(
-      false, "static_quant_epilogue requires AVX512/powerpc64/AArch64 support.")
-}
-
-template <typename scalar_t>
-void dynamic_quant_epilogue(const float* input, scalar_t* output,
-                            const float* a_scale, const float* b_scale,
-                            const int32_t* azp, const int32_t* azp_with_adj,
-                            const scalar_t* bias, const int num_tokens,
-                            const int hidden_size) {
-  TORCH_CHECK(
-      false,
-      "dynamic_quant_epilogue requires AVX512/powerpc64/AArch64 support.")
-}
-#endif
-}  // namespace
-
-void int8_scaled_mm(torch::Tensor& c,               // [M, OC], row-major
-                    const torch::Tensor& a,         // [M, IC], row-major
-                    const torch::Tensor& b,         // [IC, OC], column-major
-                    const torch::Tensor& a_scales,  // [1] or [M]
-                    const torch::Tensor& b_scales,  // [1] or [OC]
-                    const std::optional<torch::Tensor>& bias  // [OC]
-) {
-  CPU_KERNEL_GUARD_IN(cutlass_scaled_mm)
-  // Checks for conformality
-  TORCH_CHECK(a.dtype() == torch::kInt8 && b.dtype() == torch::kInt8,
-              "int8_scaled_mm only supports INT8 inputs.")
-  TORCH_CHECK(a.dim() == 2 && b.dim() == 2 && c.dim() == 2);
-  TORCH_CHECK(c.size(0) == a.size(0) && a.size(1) == b.size(0) &&
-              b.size(1) == c.size(1));
-  TORCH_CHECK(a_scales.numel() == 1 || a_scales.numel() == a.size(0));
-  TORCH_CHECK(b_scales.numel() == 1 || b_scales.numel() == b.size(1));
-
-  // Check for strides and alignment
-  TORCH_CHECK(a.stride(1) == 1 && c.stride(1) == 1);  // Row-major
-  TORCH_CHECK(b.stride(0) == 1);                      // Column-major
-  TORCH_CHECK(c.stride(0) % 16 == 0 &&
-              b.stride(1) % 16 == 0);  // 16 Byte Alignment
-  TORCH_CHECK(a_scales.is_contiguous() && b_scales.is_contiguous());
-
-  if (bias) {
-    TORCH_CHECK(bias->numel() == b.size(1) && bias->is_contiguous() &&
-                bias->dim() == 1);
-  }
-
-  VLLM_DISPATCH_FLOATING_TYPES(c.scalar_type(), "int8_scaled_mm", [&] {
-    if (a_scales.numel() != 1) {
-      // per-token
-      // Note: oneDNN doesn't support per-token activation quantization
-      // Ideally we want to fuse the GEMM and the scale procedure with oneDNN
-      // JIT, the intermediate data is cached in registers or L1. But for now
-      // the oneDNN GEMM code generation only supports two quantization
-      // patterns: per-tensor or per-output-channel of weight.
-      // So we have to apply the per-token scale with a 'epilogue'. In C=s_a *
-      // s_b * (A@B) + bias, the C_inter = s_b * (A@B) is computed by oneDNN
-      // GEMM, then the per-token scale (and bias) is applied with the epilogue
-      // C=s_a * C_inter + bias.
-      torch::Tensor tmp_fp32_out =
-          torch::empty_like(c, ::at::ScalarType::Float);
-      // Compute C_inter=s_b * (A@B)
-      DNNLPrimitiveHelper<true>::gemm_s8s8_jit<float, void>(
-          a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
-          tmp_fp32_out.data_ptr<float>(), nullptr, a.size(0), b.size(1),
-          a.size(1), nullptr, b_scales.data_ptr<float>(), 0, b_scales.numel());
-      if (bias.has_value()) {
-        // Compute C=s_a * C_inter + bias
-        dynamic_quant_epilogue<false, true, true>(
-            tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-            a_scales.data_ptr<float>(), nullptr, nullptr, nullptr,
-            bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
-      } else {
-        // Compute C=s_a * C_inter
-        dynamic_quant_epilogue<false, true, false, scalar_t>(
-            tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-            a_scales.data_ptr<float>(), nullptr, nullptr, nullptr, nullptr,
-            c.size(0), c.size(1));
-      }
-    } else {
-      // per-tensor
-      if (bias.has_value()) {
-        // Compute C=s_a * s_b * (A@B) + bias
-        DNNLPrimitiveHelper<false>::gemm_s8s8_jit(
-            a.data_ptr<int8_t>(), b.data_ptr<int8_t>(), c.data_ptr<scalar_t>(),
-            bias->data_ptr<scalar_t>(), a.size(0), b.size(1), a.size(1),
-            a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-            a_scales.numel(), b_scales.numel());
-      } else {
-        // Compute C=s_a * s_b * (A@B)
-        DNNLPrimitiveHelper<false>::gemm_s8s8_jit<scalar_t, void>(
-            a.data_ptr<int8_t>(), b.data_ptr<int8_t>(), c.data_ptr<scalar_t>(),
-            nullptr, a.size(0), b.size(1), a.size(1),
-            a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-            a_scales.numel(), b_scales.numel());
-      }
-    }
-  });
-}
-
-void int8_scaled_mm_azp(torch::Tensor& c,        // [M, OC], row-major
-                        const torch::Tensor& a,  // [M, IC], row-major
-                        const torch::Tensor& b,  // [IC, OC], column-major
-                        const torch::Tensor& a_scales,            // [1] or [M]
-                        const torch::Tensor& b_scales,            // [1] or [OC]
-                        const torch::Tensor& azp_adj,             // [OC]
-                        const std::optional<torch::Tensor>& azp,  // [1] or [M]
-                        const std::optional<torch::Tensor>& bias  // [OC]
-) {
-  CPU_KERNEL_GUARD_IN(cutlass_scaled_mm_azp)
-  // Checks for conformality
-  TORCH_CHECK(a.dtype() == torch::kInt8 && b.dtype() == torch::kInt8,
-              "int8_scaled_mm_azp only supports INT8 inputs.")
-  TORCH_CHECK(a.dim() == 2 && b.dim() == 2 && c.dim() == 2);
-  TORCH_CHECK(c.size(0) == a.size(0) && a.size(1) == b.size(0) &&
-              b.size(1) == c.size(1));
-  TORCH_CHECK(a_scales.numel() == 1 || a_scales.numel() == a.size(0));
-  TORCH_CHECK(b_scales.numel() == 1 || b_scales.numel() == b.size(1));
-
-  // Check for strides and alignment
-  TORCH_CHECK(a.stride(1) == 1 && c.stride(1) == 1);  // Row-major
-  TORCH_CHECK(b.stride(0) == 1);                      // Column-major
-  TORCH_CHECK(c.stride(0) % 16 == 0 &&
-              b.stride(1) % 16 == 0);  // 16 Byte Alignment
-  TORCH_CHECK(a_scales.is_contiguous() && b_scales.is_contiguous());
-
-  if (bias) {
-    TORCH_CHECK(bias->numel() == b.size(1) && bias->is_contiguous());
-  }
-  if (azp) {
-    TORCH_CHECK(azp->numel() == a.size(0) && azp->is_contiguous());
-  }
-  TORCH_CHECK(azp_adj.numel() == b.size(1) && azp_adj.is_contiguous());
-
-  // azp & bias types
-  TORCH_CHECK(azp_adj.dtype() == torch::kInt32);
-  TORCH_CHECK(!azp || azp->dtype() == torch::kInt32);
-  TORCH_CHECK(!bias || bias->dtype() == c.dtype(),
-              "currently bias dtype must match output dtype ", c.dtype());
-
-  VLLM_DISPATCH_FLOATING_TYPES(c.scalar_type(), "int8_scaled_mm_azp", [&] {
-    torch::Tensor tmp_fp32_out = torch::empty_like(c, ::at::ScalarType::Float);
-    if (a_scales.numel() != 1) {
-      // per-token
-      // Note: oneDNN doesn't support per-token activation quantization
-      // Compute C_inter=s_b * (A@B)
-      DNNLPrimitiveHelper<true>::gemm_s8s8_jit<float, void>(
-          a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
-          tmp_fp32_out.data_ptr<float>(), nullptr, a.size(0), b.size(1),
-          a.size(1), nullptr, b_scales.data_ptr<float>(), 0, b_scales.numel());
-      if (bias.has_value()) {
-        // Compute C=s_a * C_inter - s_a * s_b * azp * azp_adj + bias
-        if (b_scales.numel() != 1) {
-          // Per-Channel
-          dynamic_quant_epilogue<true, true, true>(
-              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-              a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-              azp->data_ptr<int32_t>(), azp_adj.data_ptr<int32_t>(),
-              bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
-        } else {
-          // Per-Tensor
-          dynamic_quant_epilogue<true, false, true>(
-              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-              a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-              azp->data_ptr<int32_t>(), azp_adj.data_ptr<int32_t>(),
-              bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
-        }
-      } else {
-        // Compute C=s_a * C_inter - s_a * s_b * azp * azp_adj
-        if (b_scales.numel() != 1) {
-          // Per-Channel
-          dynamic_quant_epilogue<true, true, false, scalar_t>(
-              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-              a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-              azp->data_ptr<int32_t>(), azp_adj.data_ptr<int32_t>(), nullptr,
-              c.size(0), c.size(1));
-        } else {
-          // Per-Tensor
-          dynamic_quant_epilogue<true, false, false, scalar_t>(
-              tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-              a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-              azp->data_ptr<int32_t>(), azp_adj.data_ptr<int32_t>(), nullptr,
-              c.size(0), c.size(1));
-        }
-      }
-    } else {
-      // per-tensor
-      if (bias.has_value()) {
-        // Compute C_inter=s_a * s_b * (A@B) + bias
-        DNNLPrimitiveHelper<false>::gemm_s8s8_jit(
-            a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
-            tmp_fp32_out.data_ptr<float>(), bias->data_ptr<scalar_t>(),
-            a.size(0), b.size(1), a.size(1), a_scales.data_ptr<float>(),
-            b_scales.data_ptr<float>(), a_scales.numel(), b_scales.numel());
-      } else {
-        // Compute C_inter=s_a * s_b * (A@B)
-        DNNLPrimitiveHelper<false>::gemm_s8s8_jit<float, void>(
-            a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
-            tmp_fp32_out.data_ptr<float>(), nullptr, a.size(0), b.size(1),
-            a.size(1), a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-            a_scales.numel(), b_scales.numel());
-      }
-
-      // Compute C=C_inter - s_a * s_b * azp_adj
-      if (b_scales.numel() != 1) {
-        // Per-Channel
-        static_quant_epilogue<true>(
-            tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-            *a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-            azp_adj.data_ptr<int32_t>(), a.size(0), b.size(1));
-      } else {
-        // Per-Tensor
-        static_quant_epilogue<false>(
-            tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-            *a_scales.data_ptr<float>(), b_scales.data_ptr<float>(),
-            azp_adj.data_ptr<int32_t>(), a.size(0), b.size(1));
-      }
-    }
-  });
-}
-
-// static-per-tensor quantization.
-void static_scaled_int8_quant(torch::Tensor& out,          // [..., hidden_size]
-                              const torch::Tensor& input,  // [..., hidden_size]
-                              const torch::Tensor& scale,
-                              std::optional<torch::Tensor> const& azp) {
-  CPU_KERNEL_GUARD_IN(static_scaled_int8_quant)
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(out.is_contiguous());
-  TORCH_CHECK(scale.numel() == 1);
-  TORCH_CHECK(!azp.has_value() || azp->numel() == 1);
-
-  const int hidden_size = input.size(-1);
-  const int num_tokens = input.numel() / hidden_size;
-  VLLM_DISPATCH_FLOATING_TYPES(
-      input.scalar_type(), "static_scaled_int8_quant_impl", [&] {
-        if (azp.has_value()) {
-          static_scaled_int8_quant_impl<true>(
-              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
-              scale.data_ptr<float>(), azp->data_ptr<int32_t>(), num_tokens,
-              hidden_size);
-        } else {
-          static_scaled_int8_quant_impl<false>(
-              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
-              scale.data_ptr<float>(), nullptr, num_tokens, hidden_size);
-        }
-      });
-}
-
-// dynamic-per-token quantization.
-void dynamic_scaled_int8_quant(
-    torch::Tensor& out,          // [..., hidden_size]
-    const torch::Tensor& input,  // [..., hidden_size]
-    torch::Tensor& scale,        // [..., 1]
-    std::optional<torch::Tensor> const& azp) {
-  CPU_KERNEL_GUARD_IN(dynamic_scaled_int8_quant)
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(out.is_contiguous());
-
-  int const hidden_size = input.size(-1);
-  int const num_tokens = input.numel() / hidden_size;
-  VLLM_DISPATCH_FLOATING_TYPES(
-      input.scalar_type(), "dynamic_scaled_int8_quant_impl", [&] {
-        if (azp.has_value()) {
-          dynamic_scaled_int8_quant_impl<true>(
-              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
-              scale.data_ptr<float>(), azp->data_ptr<int32_t>(), num_tokens,
-              hidden_size);
-        } else {
-          dynamic_scaled_int8_quant_impl<false>(
-              input.data_ptr<scalar_t>(), out.data_ptr<int8_t>(),
-              scale.data_ptr<float>(), nullptr, num_tokens, hidden_size);
-        }
-      });
-}
-
-#if defined(__powerpc64__)
-void int8_scaled_mm_ppc64le(torch::Tensor& c,        // [M, OC], row-major
-                            const torch::Tensor& a,  // [M, IC], row-major
-                            const torch::Tensor& b,  // [IC, OC], column-major
-                            const torch::Tensor& a_scales,
-                            const torch::Tensor& b_scales,
-                            const std::optional<torch::Tensor>& bias  // [OC]
-) {
-  CPU_KERNEL_GUARD_IN(cutlass_scaled_mm)
-  // Checks for conformality
-  TORCH_CHECK(a.dtype() == torch::kInt8 && b.dtype() == torch::kInt8,
-              "int8_scaled_mm_ppc64le only supports INT8 inputs.");
-  TORCH_CHECK(a.dim() == 2 && b.dim() == 2 && c.dim() == 2);
-  TORCH_CHECK(c.size(0) == a.size(0) && a.size(1) == b.size(0) &&
-              b.size(1) == c.size(1));
-  // We dont need this
-  TORCH_CHECK(a_scales.numel() == 1 || a_scales.numel() == a.size(0));
-  TORCH_CHECK(b_scales.numel() == 1 || b_scales.numel() == b.size(1));
-
-  // Check for strides and alignment
-  TORCH_CHECK(a.stride(1) == 1 && c.stride(1) == 1);  // Row-major
-  TORCH_CHECK(b.stride(0) == 1);                      // Column-major
-  TORCH_CHECK(c.stride(0) % 16 == 0 &&
-              b.stride(1) % 16 == 0);  // 16 Byte Alignment
-  TORCH_CHECK(a_scales.is_contiguous() && b_scales.is_contiguous());
-
-  if (bias) {
-    TORCH_CHECK(bias->numel() == b.size(1) && bias->is_contiguous() &&
-                bias->dim() == 1);
-  }
-  VLLM_DISPATCH_FLOATING_TYPES(c.scalar_type(), "int8_scaled_mm_ppc64le", [&] {
-    torch::Tensor tmp_fp32_out = torch::empty_like(c, ::at::ScalarType::Float);
-    // Compute C_inter=s_b * (A@B)
-    DNNLPrimitiveHelper<true>::gemm_s8s8_jit<float, void>(
-        a.data_ptr<int8_t>(), b.data_ptr<int8_t>(),
-        tmp_fp32_out.data_ptr<float>(), nullptr, a.size(0), b.size(1),
-        a.size(1), nullptr, b_scales.data_ptr<float>(), 0, b_scales.numel());
-    if (bias.has_value()) {
-      // Compute C=s_a * C_inter + bias
-      dynamic_quant_epilogue<false, true, true>(
-          tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-          a_scales.data_ptr<float>(), nullptr, nullptr, nullptr,
-          bias->data_ptr<scalar_t>(), c.size(0), c.size(1));
-    } else {
-      // Compute C=s_a * C_inter
-      dynamic_quant_epilogue<false, true, false, scalar_t>(
-          tmp_fp32_out.data_ptr<float>(), c.data_ptr<scalar_t>(),
-          a_scales.data_ptr<float>(), nullptr, nullptr, nullptr, nullptr,
-          c.size(0), c.size(1));
-    }
-  });
-}
-
-#endif

csrc/cpu/torch_bindings.cpp

@@ -6,25 +6,20 @@
 
 std::string init_cpu_threads_env(const std::string& cpu_ids);
 
-void int8_scaled_mm(torch::Tensor& c, const torch::Tensor& a,
-                    const torch::Tensor& b, const torch::Tensor& a_scales,
-                    const torch::Tensor& b_scales,
-                    const std::optional<torch::Tensor>& bias);
+void release_dnnl_matmul_handler(int64_t handler);
 
-void int8_scaled_mm_azp(torch::Tensor& c, const torch::Tensor& a,
-                        const torch::Tensor& b, const torch::Tensor& a_scales,
+int64_t create_onednn_scaled_mm_handler(const torch::Tensor& b,
                                         const torch::Tensor& b_scales,
-                        const torch::Tensor& azp_adj,
-                        const std::optional<torch::Tensor>& azp,
-                        const std::optional<torch::Tensor>& bias);
+                                        at::ScalarType output_type,
+                                        bool dynamic_act_quant, bool use_azp,
+                                        int64_t primitive_cache_size);
 
-#if defined(__powerpc64__)
-void int8_scaled_mm_ppc64le(torch::Tensor& c, const torch::Tensor& a,
-                            const torch::Tensor& b,
+void onednn_scaled_mm(torch::Tensor& c, const torch::Tensor& a,
                       const torch::Tensor& a_scales,
-                            const torch::Tensor& b_scales,
-                            const std::optional<torch::Tensor>& bias);
-#endif
+                      const std::optional<torch::Tensor>& azp,
+                      const std::optional<torch::Tensor>& azp_adj,
+                      const std::optional<torch::Tensor>& bias,
+                      int64_t handler);
 
 void mla_decode_kvcache(torch::Tensor& out, torch::Tensor& query,
                         torch::Tensor& kv_cache, double scale,
@@ -151,8 +146,25 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("rotary_embedding", torch::kCPU, &rotary_embedding);
 
   // Quantization
-#if defined(__AVX512F__) || (defined(__aarch64__) && !defined(__APPLE__))
+#if defined(__AVX512F__) || (defined(__aarch64__) && !defined(__APPLE__)) || \
+    defined(__powerpc64__)
   at::Tag stride_tag = at::Tag::needs_fixed_stride_order;
+  // Helper function to release oneDNN handlers
+  ops.def("release_dnnl_matmul_handler(int handler) -> ()",
+          &release_dnnl_matmul_handler);
+
+  // Create oneDNN W8A8 handler
+  ops.def(
+      "create_onednn_scaled_mm_handler(Tensor b, Tensor b_scales, ScalarType "
+      "output_type, bool dynamic_act_quant, bool use_azp, int "
+      "primitive_cache_size) -> int",
+      &create_onednn_scaled_mm_handler);
+
+  // oneDNN scaled_mm for W8A8 with static per-tensor activation quantization
+  ops.def(
+      "onednn_scaled_mm(Tensor! c, Tensor a, Tensor a_scales, Tensor? azp, "
+      "Tensor? azp_adj, Tensor? bias, int handler) -> ()");
+  ops.impl("onednn_scaled_mm", torch::kCPU, &onednn_scaled_mm);
 
   // Compute int8 quantized tensor for given scaling factor.
   ops.def(
@@ -168,50 +180,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       {stride_tag});
   ops.impl("dynamic_scaled_int8_quant", torch::kCPU,
            &dynamic_scaled_int8_quant);
-  // W8A8 GEMM, supporting symmetric per-tensor or per-row/column
-  // quantization.
-  ops.def(
-      "cutlass_scaled_mm(Tensor! out, Tensor a,"
-      "                  Tensor b, Tensor a_scales,"
-      "                  Tensor b_scales, Tensor? bias) -> ()",
-      {stride_tag});
-  ops.impl("cutlass_scaled_mm", torch::kCPU, &int8_scaled_mm);
-  // w8a8 GEMM, supporting asymmetric per-tensor or per-row/column
-  // quantization.
-  ops.def(
-      "cutlass_scaled_mm_azp(Tensor! out, Tensor a,"
-      "                  Tensor b, Tensor a_scales,"
-      "                  Tensor b_scales, Tensor azp_adj,"
-      "                  Tensor? azp, Tensor? bias) -> ()",
-      {stride_tag});
-  ops.impl("cutlass_scaled_mm_azp", torch::kCPU, &int8_scaled_mm_azp);
-#elif defined(__powerpc64__)
-  // Compute int8 quantized tensor for given scaling factor.
-  ops.def(
-      "static_scaled_int8_quant(Tensor! out, Tensor input, Tensor scale,"
-      "Tensor? azp) -> ()");
-  ops.impl("static_scaled_int8_quant", torch::kCPU, &static_scaled_int8_quant);
-
-  // Compute int8 quantized tensor and scaling factor
-  ops.def(
-      "dynamic_scaled_int8_quant(Tensor! out, Tensor input, Tensor! scale, "
-      "Tensor!? azp) -> ()");
-  ops.impl("dynamic_scaled_int8_quant", torch::kCPU,
-           &dynamic_scaled_int8_quant);
-  // W8A8 GEMM, supporting symmetric quantization.
-  ops.def(
-      "cutlass_scaled_mm(Tensor! out, Tensor a,"
-      "                  Tensor b, Tensor a_scales,"
-      "                  Tensor b_scales, Tensor? bias) -> ()");
-  ops.impl("cutlass_scaled_mm", torch::kCPU, &int8_scaled_mm_ppc64le);
-  // w8a8 GEMM, supporting asymmetric per-tensor or per-row/column
-  // quantization.
-  ops.def(
-      "cutlass_scaled_mm_azp(Tensor! out, Tensor a,"
-      "                  Tensor b, Tensor a_scales,"
-      "                  Tensor b_scales, Tensor azp_adj,"
-      "                  Tensor? azp, Tensor? bias) -> ()");
-  ops.impl("cutlass_scaled_mm_azp", torch::kCPU, &int8_scaled_mm_azp);
 #endif
 
 // SHM CCL

tests/kernels/test_onednn.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Integration tests for FlexAttention backend vs default backend"""
+
+from typing import Optional
+
+import pytest
+import torch
+
+from tests.kernels.utils import to_int8
+from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
+
+if not current_platform.is_cpu():
+    pytest.skip("skipping CPU-only tests", allow_module_level=True)
+
+NK_FACTORS = [
+    (256, 128),
+    (4096, 4096),
+    (16384, 4096),
+    (1023, 491),
+    (1001, 15),
+]
+M_FACTORS = [
+    (16, 1, 32, 128, 64),
+    (1, 17, 1, 31, 17),
+]
+CACHE_SIZES = [2]
+DTYPE = [torch.bfloat16]
+
+
+def rand_int8(shape: tuple, device: str = "cpu"):
+    return to_int8(torch.rand(shape, device=device) * 255 - 128)
+
+
+def ref_int8_scaled_mm(
+    a: torch.Tensor,
+    b: torch.Tensor,
+    scale_a: torch.Tensor,
+    scale_b: torch.Tensor,
+    azp: Optional[torch.Tensor],
+    bias: Optional[torch.Tensor],
+    output_type: torch.dtype,
+):
+    if azp is not None:
+        a = a.to(dtype=torch.float32) - azp.to(dtype=torch.float32)
+    output = torch.mm((scale_a * a.to(dtype=torch.float32)),
+                      (scale_b * b.to(dtype=torch.float32)))
+    if bias is not None:
+        output += bias.float()
+
+    return output.to(dtype=output_type)
+
+
+def onednn_int8_gemm_test_helper(primitive_cache_size: int,
+                                 m: int,
+                                 n: int,
+                                 k: int,
+                                 per_tensor_a_quant: bool,
+                                 per_tensor_b_quant: bool,
+                                 use_azp: bool,
+                                 use_bias: bool,
+                                 out_dtype: torch.dtype = torch.bfloat16,
+                                 device: str = "cpu"):
+    # Test for a oneDNN kernel with per-tensor / per-token activation
+    # quantization and per-tensor / per-output channel weight quantization.
+    a = to_int8(torch.randn((m, k), device=device) * 5)
+    b = to_int8(torch.randn((n, k), device=device).t() * 5)
+
+    a_scales_shape = (1, 1) if per_tensor_a_quant else (m, 1)
+    b_scales_shape = (1, 1) if per_tensor_b_quant else (1, n)
+
+    scale_a = (torch.randn(a_scales_shape, device=device, dtype=torch.float32))
+    scale_b = (torch.randn(b_scales_shape, device=device, dtype=torch.float32))
+
+    if use_azp:
+        azp = torch.rand(a_scales_shape, dtype=torch.float32) * 10 + 1.5
+        azp = (azp / scale_a).round().to(dtype=torch.int32)
+        azp_adj = scale_b * b.sum(dim=0, keepdim=True, dtype=torch.float32)
+    else:
+        azp = None
+        azp_adj = None
+
+    if use_bias:
+        bias = torch.rand((n, ), device=device, dtype=out_dtype) * 10
+    else:
+        bias = None
+
+    handler = ops.create_onednn_scaled_mm(
+        b,
+        scale_b,
+        out_dtype,
+        not per_tensor_a_quant,
+        use_azp,
+        primitive_cache_size,
+    )
+
+    out = torch.zeros((m, n), dtype=out_dtype)
+    ops.onednn_scaled_mm(handler, a, out, scale_a, azp, azp_adj, bias)
+    baseline = ref_int8_scaled_mm(a, b, scale_a, scale_b, azp, bias, out_dtype)
+
+    torch.testing.assert_close(out, baseline, rtol=1e-1, atol=1e0)
+
+    if use_bias:
+        # To test runtime bias setting
+        out = torch.zeros((m, n), dtype=out_dtype)
+        ops.onednn_scaled_mm(handler, a, out, scale_a, azp, azp_adj, None)
+        baseline = ref_int8_scaled_mm(a, b, scale_a, scale_b, azp, None,
+                                      out_dtype)
+
+        torch.testing.assert_close(out, baseline, rtol=1e-1, atol=1e0)
+
+
+@pytest.mark.parametrize("n,k", NK_FACTORS)
+@pytest.mark.parametrize("m_list", M_FACTORS)
+@pytest.mark.parametrize("per_tensor_a_scale", [True, False])
+@pytest.mark.parametrize("per_tensor_b_scale", [True, False])
+@pytest.mark.parametrize("use_bias", [True, False])
+@pytest.mark.parametrize("use_azp", [True, False])
+@pytest.mark.parametrize("output_type", DTYPE)
+@pytest.mark.parametrize("primitive_cache_size", CACHE_SIZES)
+def test_onednn_int8_scaled_gemm(
+    n: int,
+    k: int,
+    m_list: tuple[int],
+    per_tensor_a_scale: bool,
+    per_tensor_b_scale: bool,
+    use_bias: bool,
+    use_azp: bool,
+    output_type: torch.dtype,
+    primitive_cache_size: int,
+):
+    for m in m_list:
+        onednn_int8_gemm_test_helper(
+            primitive_cache_size=primitive_cache_size,
+            m=m,
+            n=n,
+            k=k,
+            per_tensor_a_quant=per_tensor_a_scale,
+            per_tensor_b_quant=per_tensor_b_scale,
+            use_bias=use_bias,
+            use_azp=use_azp,
+            out_dtype=output_type,
+        )

vllm/_custom_ops.py

@@ -1827,3 +1827,86 @@ if hasattr(torch.ops._C, "int8_scaled_mm_with_quant"):
         M = mat1.size(0)
         N = mat2.size(0)
         return torch.empty((M, N), dtype=out_dtype)
+
+
+class CPUDNNLGEMMHandler:
+
+    def __init__(self) -> None:
+        self.handler: Optional[int] = None
+        self.n = -1
+        self.k = -1
+
+    def __del__(self):
+        if self.handler is not None:
+            torch.ops._C.release_dnnl_matmul_handler(self.handler)
+
+
+def create_onednn_scaled_mm(
+    weight: torch.Tensor,  # [K, N]
+    weight_scales: torch.Tensor,
+    output_type: torch.dtype,
+    dynamic_quant: bool,
+    use_azp: bool,
+    primitive_cache_size: int = 128,
+) -> CPUDNNLGEMMHandler:
+    handler = CPUDNNLGEMMHandler()
+    handler.k, handler.n = weight.size()
+    handler.handler = torch.ops._C.create_onednn_scaled_mm_handler(
+        weight, weight_scales, output_type, dynamic_quant, use_azp,
+        primitive_cache_size)
+    return handler
+
+
+def onednn_scaled_int8_quant(input: torch.Tensor,
+                             scale: Optional[torch.Tensor] = None,
+                             azp: Optional[torch.Tensor] = None,
+                             symmetric: bool = True):
+    """
+    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
+
+    Args:
+        input: The input tensor to be quantized to int8.
+        scale: Optional scaling factor for the int8 quantization.
+            When not provided, we invoke dynamic-per-token quantization.
+        azp: Optional zero-point for the int8 quantization.
+            Must be provided for asymmetric quantization if `scale` is provided.
+        symmetric: Whether to use symmetric quantization (scale only, azp ignored).
+
+    Returns:
+      tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
+    """
+    output = torch.empty_like(input, dtype=torch.int8)
+    token_num = input.numel() // input.shape[-1]
+    input = input.view((token_num, input.shape[-1]))
+    if scale is not None:
+        # static-per-tensor quantization.
+        assert symmetric == (
+            azp
+            is None), "azp must only be provided for asymmetric quantization."
+        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
+        return output, scale, azp
+
+    # dynamic-per-token quantization.
+    input_scales = torch.empty((token_num, 1),
+                               device=input.device,
+                               dtype=torch.float32)
+    input_azp = None if symmetric else torch.empty_like(input_scales,
+                                                        dtype=torch.int32)
+    torch.ops._C.dynamic_scaled_int8_quant(output, input, input_scales,
+                                           input_azp)
+    return output, input_scales, input_azp
+
+
+def onednn_scaled_mm(
+    dnnl_handler: CPUDNNLGEMMHandler,
+    x: torch.Tensor,
+    output: torch.Tensor,
+    input_scale: Optional[torch.Tensor],
+    input_zp: Optional[torch.Tensor],
+    input_zp_adj: Optional[torch.Tensor],
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+    torch.ops._C.onednn_scaled_mm(output, x, input_scale, input_zp,
+                                  input_zp_adj, bias, dnnl_handler.handler)
+
+    return output

vllm/model_executor/layers/fused_moe/layer.py

@@ -360,10 +360,15 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         elif current_platform.is_cpu():
             if current_platform.get_cpu_architecture() == CpuArchEnum.X86:
                 from vllm.model_executor.layers.fused_moe import cpu_fused_moe
-                dtype = layer.w13_weight.dtype
+                from vllm.model_executor.layers.utils import (
+                    check_cpu_sgl_kernel)
+                dtype_w13 = layer.w13_weight.dtype
+                _, n_w13, k_w13 = layer.w13_weight.size()
+                dtype_w2 = layer.w2_weight.dtype
+                _, n_w2, k_w2 = layer.w2_weight.size()
                 if (envs.VLLM_CPU_SGL_KERNEL
-                        and torch._C._cpu._is_amx_tile_supported()
-                        and dtype == torch.bfloat16):
+                        and check_cpu_sgl_kernel(n_w13, k_w13, dtype_w13)
+                        and check_cpu_sgl_kernel(n_w2, k_w2, dtype_w2)):
                     packed_w13_weight = torch.ops._C.convert_weight_packed(
                         layer.w13_weight)
                     assert packed_w13_weight.size() == layer.w13_weight.size()

vllm/model_executor/layers/linear.py

@@ -199,11 +199,10 @@ class UnquantizedLinearMethod(LinearMethodBase):
 
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
         if current_platform.is_cpu() and envs.VLLM_CPU_SGL_KERNEL:
+            from vllm.model_executor.layers.utils import check_cpu_sgl_kernel
             N, K = layer.weight.size()
             dtype = layer.weight.dtype
-            if (torch._C._cpu._is_amx_tile_supported()
-                    and dtype == torch.bfloat16 and N % 32 == 0
-                    and K % 32 == 0):
+            if check_cpu_sgl_kernel(N, K, dtype):
                 packed_weight = torch.ops._C.convert_weight_packed(
                     layer.weight)
                 assert packed_weight.size() == layer.weight.size()
@@ -215,7 +214,8 @@ class UnquantizedLinearMethod(LinearMethodBase):
             else:
                 logger.warning(
                     "CPU SGL kernels require Intel AMX support,"
-                    " bfloat16 weight, IC and OC are divisible by 32.")
+                    " bf16/fp16/int8 weight, IC and OC are divisible by "
+                    "32 and 16.")
                 layer.use_cpu_sgl = False
 
     def apply(self,

vllm/model_executor/layers/quantization/kernels/scaled_mm/__init__.py

@@ -6,6 +6,8 @@ from typing import Optional
 
 from vllm.model_executor.layers.quantization.kernels.scaled_mm.aiter import (
     AiterScaledMMLinearKernel)
+from vllm.model_executor.layers.quantization.kernels.scaled_mm.cpu import (
+    CPUScaledMMLinearKernel)
 from vllm.model_executor.layers.quantization.kernels.scaled_mm.cutlass import (
     CutlassScaledMMLinearKernel)
 from vllm.model_executor.layers.quantization.kernels.scaled_mm.ScaledMMLinearKernel import (  # noqa: E501
@@ -18,7 +20,7 @@ from vllm.platforms import PlatformEnum, current_platform
 
 # in priority/performance order (when available)
 _POSSIBLE_KERNELS: dict[PlatformEnum, list[type[ScaledMMLinearKernel]]] = {
-    PlatformEnum.CPU: [CutlassScaledMMLinearKernel],
+    PlatformEnum.CPU: [CPUScaledMMLinearKernel],
     PlatformEnum.CUDA: [CutlassScaledMMLinearKernel],
     PlatformEnum.ROCM: [AiterScaledMMLinearKernel, TritonScaledMMLinearKernel],
     PlatformEnum.TPU: [XLAScaledMMLinearKernel],

vllm/model_executor/layers/quantization/kernels/scaled_mm/cpu.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Optional
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm import envs
+from vllm.model_executor.layers.quantization.utils import replace_parameter
+from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
+    convert_to_channelwise)
+from vllm.model_executor.layers.utils import check_cpu_sgl_kernel
+from vllm.platforms import current_platform
+from vllm.platforms.interface import CpuArchEnum
+
+from .ScaledMMLinearKernel import (ScaledMMLinearKernel,
+                                   ScaledMMLinearLayerConfig)
+
+
+class CPUScaledMMLinearKernel(ScaledMMLinearKernel):
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 75
+
+    @classmethod
+    def can_implement(
+            cls, c: ScaledMMLinearLayerConfig) -> tuple[bool, Optional[str]]:
+        if not current_platform.is_cpu():
+            return False, "CPUScaledMM requires running on CPU."
+
+        return True, None
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        weight = getattr(layer, self.w_q_name)
+        dtype = weight.dtype
+        N, K = weight.size()
+        if (current_platform.get_cpu_architecture() == CpuArchEnum.X86
+                and envs.VLLM_CPU_SGL_KERNEL and self.config.input_symmetric
+                and check_cpu_sgl_kernel(N, K, dtype)):
+            self.linear_method = self._apply_weights_sgl
+            self.process_weights_for_sgl(layer)
+        else:
+            self.linear_method = self._apply_weights_onednn
+            self.process_weights_for_onednn(layer)
+
+    def process_weights_for_onednn(self, layer: torch.nn.Module) -> None:
+        # WEIGHT
+        # Transpose to [K, N] for convenience
+        weight = getattr(layer, self.w_q_name)
+        replace_parameter(
+            layer, self.w_q_name,
+            torch.nn.Parameter(weight.t().data, requires_grad=False))
+
+        # WEIGHT SCALE
+        # oneDNN kernels support only per-tensor and per-channel.
+        # If we have a fused module (QKV, MLP) with per tensor scales (thus N
+        # scales being passed to the kernel), convert to the per-channel case.
+        is_fused_module = len(layer.logical_widths) > 1
+        weight_scale = getattr(layer, self.w_s_name)
+        if is_fused_module and not self.config.is_channelwise:
+            weight_scale = convert_to_channelwise(weight_scale,
+                                                  layer.logical_widths)
+        replace_parameter(
+            layer, self.w_s_name,
+            torch.nn.Parameter(weight_scale.data, requires_grad=False))
+
+        # INPUT SCALE
+        if self.config.is_static_input_scheme:
+            input_scale = getattr(layer, self.i_s_name)
+
+            if self.config.input_symmetric:
+                replace_parameter(
+                    layer, self.i_s_name,
+                    torch.nn.Parameter(input_scale.max(), requires_grad=False))
+                setattr(layer, self.i_zp_name, None)
+            else:
+                input_zero_point = getattr(layer, self.i_zp_name)
+
+                # reconstruct the ranges
+                int8_traits = torch.iinfo(torch.int8)
+                azps = input_zero_point.to(dtype=torch.int32)
+                range_max = (input_scale * (int8_traits.max - azps)).max()
+                range_min = (input_scale * (int8_traits.min - azps)).min()
+
+                scale = (range_max - range_min) / (int8_traits.max -
+                                                   int8_traits.min)
+                replace_parameter(
+                    layer, self.i_s_name,
+                    torch.nn.Parameter(scale, requires_grad=False))
+
+                azp = (int8_traits.min -
+                       range_min / scale).round().to(dtype=torch.int32)
+                replace_parameter(layer, self.i_zp_name,
+                                  torch.nn.Parameter(azp, requires_grad=False))
+
+        else:
+            setattr(layer, self.i_s_name, None)
+            setattr(layer, self.i_zp_name, None)
+
+        # Different from cutlass, oneDNN kernels only need the AZP adjustment
+        # term for dynamic quantization. And s_b should be folded into the
+        # term. Such as:
+        # s_a * s_b * [(A - zp_a)B] + bias =
+        # s_a * (s_b * AB) - s_a * s_b * zp_a * B + bias =
+        # s_a * GEMM_output - s_a * zp_a * adj + bias
+        if not (self.config.input_symmetric
+                and self.config.is_static_input_scheme):
+            weight = getattr(layer, self.w_q_name)
+            weight_scale = getattr(layer, self.w_s_name)
+            azp_adj = weight.sum(dim=0, keepdim=True, dtype=torch.float32)
+            azp_adj = azp_adj * weight_scale.squeeze()
+            setattr(layer, self.azp_adj_name,
+                    torch.nn.Parameter(azp_adj, requires_grad=False))
+        else:
+            setattr(layer, self.azp_adj_name, None)
+
+        weight = getattr(layer, self.w_q_name)
+        self.dnnl_handler = ops.create_onednn_scaled_mm(
+            weight,
+            getattr(layer, self.w_s_name),
+            torch.get_default_dtype(),
+            getattr(layer, self.i_s_name) is None,
+            not self.config.input_symmetric,
+            32,
+        )
+        # weight is prepacked and maintained by the dnnl_handler,
+        # release the original weight
+        setattr(layer, self.w_q_name, None)
+        del weight
+
+    def process_weights_for_sgl(self, layer: torch.nn.Module) -> None:
+        # WEIGHT
+        weight = getattr(layer, self.w_q_name)
+        packed_weight = torch.ops._C.convert_weight_packed(weight)
+        replace_parameter(
+            layer, self.w_q_name,
+            torch.nn.Parameter(packed_weight, requires_grad=False))
+
+        if layer.bias is not None:
+            bias = layer.bias
+            layer.register_parameter(
+                "bias_fp32",
+                torch.nn.Parameter(bias.float().data, requires_grad=False))
+
+        # WEIGHT SCALE
+        # CPU SGL kernels only support per-channel.
+        # For per-tensor quant, convert to the per-channel case.
+        weight_scale = getattr(layer, self.w_s_name)
+        if not self.config.is_channelwise:
+            weight_scale = convert_to_channelwise(weight_scale,
+                                                  layer.logical_widths)
+        replace_parameter(
+            layer, self.w_s_name,
+            torch.nn.Parameter(weight_scale.data, requires_grad=False))
+
+        setattr(layer, self.i_s_name, None)
+        setattr(layer, self.i_zp_name, None)
+        setattr(layer, self.azp_adj_name, None)
+
+    def apply_weights(self,
+                      layer: torch.nn.Module,
+                      x: torch.Tensor,
+                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        return self.linear_method(
+            layer,
+            x,
+            bias,
+        )
+
+    def _apply_weights_onednn(
+            self,
+            layer: torch.nn.Module,
+            x: torch.Tensor,
+            bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        w_q, w_s, i_s, i_zp, azp_adj = self._get_weight_params(layer)
+
+        # ops.scaled_int8_quant supports both dynamic and static quant:
+        # * dynamic, i_s is None and x_s computed from x.
+        # * static, i_s is scalar and x_s is i_s.
+        x_q, x_s, x_zp = ops.onednn_scaled_int8_quant(
+            x, i_s, i_zp, self.config.input_symmetric)
+
+        m = x.size(0)
+        n = self.dnnl_handler.n
+        out = torch.empty((m, n), dtype=x.dtype)
+        ops.onednn_scaled_mm(self.dnnl_handler, x_q, out, x_s, x_zp, azp_adj,
+                             bias)
+
+        return out
+
+    def _apply_weights_sgl(
+            self,
+            layer: torch.nn.Module,
+            x: torch.Tensor,
+            bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        w_q, w_s, _, _, _ = self._get_weight_params(layer)
+        return torch.ops._C.int8_scaled_mm_with_quant(
+            x,
+            w_q,
+            w_s,
+            layer.bias_fp32 if bias is not None else None,
+            x.dtype,
+            True,
+        )

vllm/model_executor/layers/quantization/kernels/scaled_mm/cutlass.py

@@ -25,8 +25,8 @@ class CutlassScaledMMLinearKernel(ScaledMMLinearKernel):
     def can_implement(
             cls, c: ScaledMMLinearLayerConfig) -> tuple[bool, Optional[str]]:
 
-        if (not current_platform.is_cuda() and not current_platform.is_cpu()):
-            return False, "CutlassScaledMM requires running on CUDA or CPU."
+        if not current_platform.is_cuda():
+            return False, "CutlassScaledMM requires running on CUDA."
 
         return True, None
 

vllm/model_executor/layers/utils.py

@@ -142,6 +142,12 @@ direct_register_custom_op(
 )
 
 
+def check_cpu_sgl_kernel(n: int, k: int, dtype: torch.dtype):
+    return (torch._C._cpu._is_amx_tile_supported()
+            and (dtype in (torch.bfloat16, torch.int8)) and k % 32 == 0
+            and n % 16 == 0)
+
+
 def cpu_unquantized_gemm(layer: torch.nn.Module,
                          x: torch.Tensor,
                          weight: torch.Tensor,