diff --git a/Dockerfile.arm b/Dockerfile.arm
new file mode 100644
index 0000000000000..093ee2209222f
--- /dev/null
+++ b/Dockerfile.arm
@@ -0,0 +1,62 @@
+# This vLLM Dockerfile is used to construct an image that can build and run vLLM on ARM CPU platform.
+
+FROM ubuntu:22.04 AS cpu-test-arm
+
+ENV CCACHE_DIR=/root/.cache/ccache
+
+ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
+
+RUN --mount=type=cache,target=/var/cache/apt \
+    apt-get update -y \
+    && apt-get install -y curl ccache git wget vim numactl gcc-12 g++-12 python3 python3-pip libtcmalloc-minimal4 libnuma-dev \
+    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
+    && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
+
+# tcmalloc provides better memory allocation efficiency, e.g., holding memory in caches to speed up access of commonly-used objects.
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install py-cpuinfo  # Use this to gather CPU info and optimize based on ARM Neoverse cores
+
+# Set LD_PRELOAD for tcmalloc on ARM
+ENV LD_PRELOAD="/usr/lib/aarch64-linux-gnu/libtcmalloc_minimal.so.4"
+
+RUN echo 'ulimit -c 0' >> ~/.bashrc
+
+WORKDIR /workspace
+
+ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
+ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
+RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,src=requirements-build.txt,target=requirements-build.txt \
+    pip install --upgrade pip && \
+    pip install -r requirements-build.txt
+
+FROM cpu-test-arm AS build
+
+WORKDIR /workspace/vllm
+
+RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,src=requirements-common.txt,target=requirements-common.txt \
+    --mount=type=bind,src=requirements-cpu.txt,target=requirements-cpu.txt \
+    pip install -v -r requirements-cpu.txt
+
+COPY . .
+ARG GIT_REPO_CHECK=0
+RUN --mount=type=bind,source=.git,target=.git \
+    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
+
+# Disabling AVX512 specific optimizations for ARM
+ARG VLLM_CPU_DISABLE_AVX512="true"
+ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
+
+RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=cache,target=/root/.cache/ccache \
+    --mount=type=bind,source=.git,target=.git \
+    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel && \
+    pip install dist/*.whl && \
+    rm -rf dist
+
+WORKDIR /workspace/
+
+RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
\ No newline at end of file
diff --git a/cmake/cpu_extension.cmake b/cmake/cpu_extension.cmake
index 426189481575b..68f7ca1af05ad 100644
--- a/cmake/cpu_extension.cmake
+++ b/cmake/cpu_extension.cmake
@@ -16,16 +16,15 @@ include_directories("${CMAKE_SOURCE_DIR}/csrc")
 #
 # Check the compile flags
 #
-if (CMAKE_SYSTEM_PROCESSOR STREQUAL "ppc64le")
-    list(APPEND CXX_COMPILE_FLAGS
-        "-fopenmp"
-        "-DVLLM_CPU_EXTENSION")
-else()
+
+if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64")
     list(APPEND CXX_COMPILE_FLAGS
-        "-fopenmp"
         "-mf16c"
-        "-DVLLM_CPU_EXTENSION")
+    )
 endif()
+list(APPEND CXX_COMPILE_FLAGS
+    "-fopenmp"
+    "-DVLLM_CPU_EXTENSION")
 
 execute_process(COMMAND cat /proc/cpuinfo
                 RESULT_VARIABLE CPUINFO_RET
@@ -59,6 +58,8 @@ find_isa(${CPUINFO} "avx2" AVX2_FOUND)
 find_isa(${CPUINFO} "avx512f" AVX512_FOUND)
 find_isa(${CPUINFO} "POWER10" POWER10_FOUND)
 find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
+find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
+find_isa(${CPUINFO} "bf16" ARM_BF16_FOUND) # Check for ARM BF16 support
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     list(APPEND CXX_COMPILE_FLAGS
@@ -78,9 +79,11 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
     else()
         message(WARNING "Disable AVX512-BF16 ISA support, no avx512_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512BF16=1.")
     endif()
+    
 elseif (AVX2_FOUND)
     list(APPEND CXX_COMPILE_FLAGS "-mavx2")
     message(WARNING "vLLM CPU backend using AVX2 ISA")
+    
 elseif (POWER9_FOUND OR POWER10_FOUND)
     message(STATUS "PowerPC detected")
     # Check for PowerPC VSX support
@@ -88,8 +91,20 @@ elseif (POWER9_FOUND OR POWER10_FOUND)
         "-mvsx"
         "-mcpu=native"
         "-mtune=native")
+
+elseif (ASIMD_FOUND)
+    message(STATUS "ARMv8 or later architecture detected")
+    if(ARM_BF16_FOUND)
+        message(STATUS "BF16 extension detected")
+        set(MARCH_FLAGS "-march=armv8.2-a+bf16+dotprod+fp16")
+        add_compile_definitions(ARM_BF16_SUPPORT)
+    else()
+        message(WARNING "BF16 functionality is not available")
+        set(MARCH_FLAGS "-march=armv8.2-a+dotprod+fp16")  
+    endif()
+    list(APPEND CXX_COMPILE_FLAGS ${MARCH_FLAGS})     
 else()
-    message(FATAL_ERROR "vLLM CPU backend requires AVX512 or AVX2 or Power9+ ISA support.")
+    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA or ARMv8 support.")
 endif()
 
 #
@@ -159,4 +174,4 @@ define_gpu_extension_target(
     WITH_SOABI
 )
 
-message(STATUS "Enabling C extension.")
+message(STATUS "Enabling C extension.")
\ No newline at end of file
diff --git a/csrc/cpu/attention.cpp b/csrc/cpu/attention.cpp
index e6c03dcb034fd..e21832ba7582f 100644
--- a/csrc/cpu/attention.cpp
+++ b/csrc/cpu/attention.cpp
@@ -51,6 +51,10 @@ struct KernelVecType<c10::BFloat16> {
   using v_load_vec_type = vec_op::BF16Vec16;
 };
 #else
+  #ifdef __aarch64__
+    #ifndef ARM_BF16_SUPPORT
+    // pass
+    #else
 template <>
 struct KernelVecType<c10::BFloat16> {
   using q_load_vec_type = vec_op::BF16Vec8;
@@ -60,6 +64,18 @@ struct KernelVecType<c10::BFloat16> {
   using qk_acc_vec_type = vec_op::FP32Vec16;
   using v_load_vec_type = vec_op::BF16Vec16;
 };
+    #endif
+  #else
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using q_load_vec_type = vec_op::BF16Vec8;
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_load_vec_type = vec_op::BF16Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+  #endif
 #endif
 
 template <typename T>
@@ -779,4 +795,4 @@ void paged_attention_v2(
                                  CALL_V2_KERNEL_LAUNCHER_BLOCK_SIZE(scalar_t);
                                  CPU_KERNEL_GUARD_OUT(paged_attention_v2_impl)
                                });
-}
+}
\ No newline at end of file
diff --git a/csrc/cpu/cpu_types.hpp b/csrc/cpu/cpu_types.hpp
index 0213be09105ed..28db0479748bf 100644
--- a/csrc/cpu/cpu_types.hpp
+++ b/csrc/cpu/cpu_types.hpp
@@ -1,4 +1,3 @@
-
 #ifndef CPU_TYPES_HPP
 #define CPU_TYPES_HPP
 
@@ -8,8 +7,11 @@
 #elif defined(__POWER9_VECTOR__)
   //ppc implementation
   #include "cpu_types_vsx.hpp"
+#elif defined(__aarch64__)
+  //arm implementation
+  #include "cpu_types_arm.hpp"
 #else
   #warning "unsupported vLLM cpu implementation"
 #endif
 
-#endif
+#endif
\ No newline at end of file
diff --git a/csrc/cpu/cpu_types_arm.hpp b/csrc/cpu/cpu_types_arm.hpp
new file mode 100644
index 0000000000000..73e0f8cb2e0fb
--- /dev/null
+++ b/csrc/cpu/cpu_types_arm.hpp
@@ -0,0 +1,515 @@
+#include <arm_neon.h>
+#include <torch/all.h> 
+#include <cmath>
+
+namespace vec_op {
+
+#ifdef ARM_BF16_SUPPORT
+  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)                                 \
+    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)                         \
+    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)                         \
+    AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)  
+#else
+  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)                                 \
+    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)                         \
+    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
+#endif
+
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...)                          \
+  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
+
+#ifndef CPU_OP_GUARD
+#define CPU_KERNEL_GUARD_IN(NAME)
+#define CPU_KERNEL_GUARD_OUT(NAME)
+#else
+#define CPU_KERNEL_GUARD_IN(NAME)                                              \
+  std::cout << #NAME << " invoked." << std::endl;
+#define CPU_KERNEL_GUARD_OUT(NAME) std::cout << #NAME << " exit." << std::endl;
+#endif
+
+#define FORCE_INLINE __attribute__((always_inline)) inline
+
+namespace {
+  template <typename T, T... indexes, typename F>
+  constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F &&f) {
+    (f(std::integral_constant<T, indexes>{}), ...);
+  };
+}; 
+
+template <typename T, T count, typename F,
+          typename = std::enable_if_t<std::is_invocable_v<F, T>>>
+constexpr void unroll_loop(F &&f) {
+  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
+}
+
+template <typename T> struct Vec {
+  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; };
+};
+
+struct FP32Vec8;
+struct FP32Vec16;
+
+struct FP16Vec8 : public Vec<FP16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+
+  float16x8_t reg;
+
+  explicit FP16Vec8(const void *ptr)
+      : reg(vld1q_f16(static_cast<const __fp16 *>(ptr))) {};
+
+  explicit FP16Vec8(const FP32Vec8 &);
+
+  void save(void *ptr) const {
+    vst1q_f16(static_cast<__fp16 *>(ptr), reg);
+  }
+};
+
+struct FP16Vec16 : public Vec<FP16Vec16> {
+    constexpr static int VEC_ELEM_NUM = 16;
+    
+    float16x8x2_t reg; 
+    
+    explicit FP16Vec16(const void *ptr) {
+        reg.val[0] = vld1q_f16(reinterpret_cast<const __fp16*>(ptr));        
+        reg.val[1] = vld1q_f16(reinterpret_cast<const __fp16*>(ptr) + 8);    
+    }
+    
+    explicit FP16Vec16(const FP32Vec16& vec);
+    
+    void save(void *ptr) const {
+        vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]);       
+        vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]);   
+    }
+    
+    void save(void *ptr, const int elem_num) const {
+        int full_blocks = elem_num / 8;   
+        int remainder = elem_num % 8;     
+        
+        if (full_blocks > 0) {
+            vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]);
+            if (full_blocks > 1) {
+                vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]);
+            }
+        }
+        
+        if (remainder > 0) {
+            float16x8_t temp = reg.val[full_blocks];
+            for (int i = 0; i < remainder; ++i) {
+                reinterpret_cast<__fp16*>(ptr)[full_blocks * 8 + i] = vgetq_lane_f16(temp, i);
+            }
+        }
+    }
+};
+
+
+#ifdef ARM_BF16_SUPPORT
+struct BF16Vec8 : public Vec<BF16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+
+  bfloat16x8_t reg;
+
+  explicit BF16Vec8(const void *ptr)
+      : reg(*reinterpret_cast<const bfloat16x8_t *>(ptr)) {};
+
+  explicit BF16Vec8(bfloat16x8_t data) : reg(data) {};
+
+  explicit BF16Vec8(const FP32Vec8 &);
+
+  explicit BF16Vec8(float32x4x2_t v) : reg(vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[0]), v.val[1])) {};  
+
+  void save(void *ptr) const { *reinterpret_cast<bfloat16x8_t *>(ptr) = reg; }
+};
+
+struct BF16Vec16 : public Vec<BF16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+
+  bfloat16x8x2_t reg;
+
+  explicit BF16Vec16(const void *ptr)
+      : reg(*reinterpret_cast<const bfloat16x8x2_t *>(ptr)) {};
+
+  explicit BF16Vec16(bfloat16x8x2_t data) : reg(data) {};
+
+  explicit BF16Vec16(const FP32Vec16 &);
+
+  explicit BF16Vec16(float32x4x4_t v) : reg({
+    vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[0]), v.val[1]),
+    vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[2]), v.val[3])
+  }){};
+
+  void save(void *ptr) const { *reinterpret_cast<bfloat16x8x2_t *>(ptr) = reg; };
+};
+
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+
+  bfloat16x8x4_t reg;
+
+  explicit BF16Vec32(const void *ptr)
+      : reg(*reinterpret_cast<const bfloat16x8x4_t *>(ptr)) {};
+
+  explicit BF16Vec32(bfloat16x8x4_t data) : reg(data) {};
+
+  explicit BF16Vec32(const BF16Vec8 &vec8_data) : reg({
+    vec8_data.reg,
+    vec8_data.reg,
+    vec8_data.reg,
+    vec8_data.reg
+  }) {};
+
+  void save(void *ptr) const { *reinterpret_cast<bfloat16x8x4_t *>(ptr) = reg; };
+};
+#endif
+
+struct FP32Vec4 : public Vec<FP32Vec4> {
+  constexpr static int VEC_ELEM_NUM = 4;
+
+  union AliasReg {
+    float32x4_t reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  float32x4_t reg;
+
+  explicit FP32Vec4(float v) : reg(vdupq_n_f32(v)) {};
+
+  explicit FP32Vec4() : reg(vdupq_n_f32(0.0f)) {};
+
+  explicit FP32Vec4(const float *ptr) : reg(vld1q_f32(ptr)) {};
+
+  explicit FP32Vec4(float32x4_t data) : reg(data) {};
+
+  explicit FP32Vec4(const FP32Vec4 &data) : reg(data.reg) {};
+};
+
+struct FP32Vec8 : public Vec<FP32Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  union AliasReg {
+    float32x4x2_t reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  float32x4x2_t reg;
+
+  explicit FP32Vec8(float v) : reg({vmovq_n_f32(v), vmovq_n_f32(v)}) {};
+
+  explicit FP32Vec8() : reg({vmovq_n_f32(0.0), vmovq_n_f32(0.0)}) {};
+
+  explicit FP32Vec8(const float *ptr) : reg({vld1q_f32(ptr), vld1q_f32(ptr + 4)}) {};
+
+  explicit FP32Vec8(float32x4x2_t data) : reg(data) {};
+
+  explicit FP32Vec8(const FP32Vec8 &data) : reg(data.reg) {};
+
+  explicit FP32Vec8(const FP16Vec8 &v) {
+        reg.val[0] = vcvt_f32_f16(vget_low_f16(v.reg));  
+        reg.val[1] = vcvt_f32_f16(vget_high_f16(v.reg)); 
+    };
+
+  explicit FP32Vec8(float16x8_t v) : reg({vcvt_f32_f16(vget_low_f16(v)), vcvt_f32_f16(vget_high_f16(v))}) {};
+
+  #ifdef ARM_BF16_SUPPORT
+
+  explicit FP32Vec8(bfloat16x8_t v) : reg({vcvtq_low_f32_bf16(v), vcvtq_high_f32_bf16(v)}) {};
+
+  explicit FP32Vec8(const BF16Vec8 &v) : reg({vcvtq_low_f32_bf16(v.reg), vcvtq_high_f32_bf16(v.reg)}) {};
+
+  #endif
+
+  float reduce_sum() const {
+    AliasReg ar;
+    ar.reg = reg;
+    float answer = 0;
+    unroll_loop<int, VEC_ELEM_NUM>([&answer, &ar](int i) { answer += ar.values[i]; });
+
+    return answer;
+  }
+
+  FP32Vec8 exp() const {
+    AliasReg ar;
+    ar.reg = reg;
+
+    float32x2_t exp_vec0 = {expf(ar.values[0]), expf(ar.values[1])};
+    float32x2_t exp_vec1 = {expf(ar.values[2]), expf(ar.values[3])};
+    float32x2_t exp_vec2 = {expf(ar.values[4]), expf(ar.values[5])};
+    float32x2_t exp_vec3 = {expf(ar.values[6]), expf(ar.values[7])};
+
+    float32x4_t result0 = vcombine_f32(exp_vec0, exp_vec1);
+    float32x4_t result1 = vcombine_f32(exp_vec2, exp_vec3);
+
+    float32x4x2_t result;
+    result.val[0] = result0;
+    result.val[1] = result1;
+
+    return FP32Vec8(result);
+  }
+
+  FP32Vec8 tanh() const {
+    AliasReg ar;
+    ar.reg = reg;
+
+    float32x2_t tanh_vec0 = {tanhf(ar.values[0]), tanhf(ar.values[1])};
+    float32x2_t tanh_vec1 = {tanhf(ar.values[2]), tanhf(ar.values[3])};
+    float32x2_t tanh_vec2 = {tanhf(ar.values[4]), tanhf(ar.values[5])};
+    float32x2_t tanh_vec3 = {tanhf(ar.values[6]), tanhf(ar.values[7])};
+
+    float32x4_t result0 = vcombine_f32(tanh_vec0, tanh_vec1);
+    float32x4_t result1 = vcombine_f32(tanh_vec2, tanh_vec3);
+
+    float32x4x2_t result;
+    result.val[0] = result0;
+    result.val[1] = result1;
+
+    return FP32Vec8(result);
+  }
+
+  FP32Vec8 er() const {
+    AliasReg ar;
+    ar.reg = reg;
+
+    float32x2_t er_vec0 = {static_cast<float32_t>(erf(ar.values[0])), static_cast<float32_t>(erf(ar.values[1]))};
+    float32x2_t er_vec1 = {static_cast<float32_t>(erf(ar.values[2])), static_cast<float32_t>(erf(ar.values[3]))};
+    float32x2_t er_vec2 = {static_cast<float32_t>(erf(ar.values[4])), static_cast<float32_t>(erf(ar.values[5]))};
+    float32x2_t er_vec3 = {static_cast<float32_t>(erf(ar.values[6])), static_cast<float32_t>(erf(ar.values[7]))};
+
+    float32x4_t result0 = vcombine_f32(er_vec0, er_vec1);
+    float32x4_t result1 = vcombine_f32(er_vec2, er_vec3);
+
+    float32x4x2_t result;
+    result.val[0] = result0;
+    result.val[1] = result1;
+
+    return FP32Vec8(result);
+  } 
+
+  FP32Vec8 operator*(const FP32Vec8 &b) const {
+    return FP32Vec8(float32x4x2_t({vmulq_f32(reg.val[0], b.reg.val[0]), vmulq_f32(reg.val[1], b.reg.val[1])}));
+  }
+
+  FP32Vec8 operator+(const FP32Vec8 &b) const {
+    return FP32Vec8(float32x4x2_t({vaddq_f32(reg.val[0], b.reg.val[0]), vaddq_f32(reg.val[1], b.reg.val[1])}));
+  }
+
+  FP32Vec8 operator-(const FP32Vec8 &b) const {
+    return FP32Vec8(float32x4x2_t({vsubq_f32(reg.val[0], b.reg.val[0]), vsubq_f32(reg.val[1], b.reg.val[1])}));
+  }
+
+  FP32Vec8 operator/(const FP32Vec8 &b) const {
+    return FP32Vec8(float32x4x2_t({vdivq_f32(reg.val[0], b.reg.val[0]), vdivq_f32(reg.val[1], b.reg.val[1])}));
+  }
+
+  void save(float *ptr) const {
+    vst1q_f32(ptr, reg.val[0]);
+    vst1q_f32(ptr + 4, reg.val[1]);
+  }
+};
+
+struct FP32Vec16 : public Vec<FP32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  union AliasReg {
+    float32x4x4_t reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  float32x4x4_t reg;
+
+  explicit FP32Vec16(float v) : reg({vmovq_n_f32(v), vmovq_n_f32(v), vmovq_n_f32(v), vmovq_n_f32(v)}) {}
+
+  explicit FP32Vec16() : reg({vmovq_n_f32(0.0), vmovq_n_f32(0.0), vmovq_n_f32(0.0), vmovq_n_f32(0.0)}) {}
+
+  explicit FP32Vec16(const float *ptr) : reg({vld1q_f32(ptr), vld1q_f32(ptr + 4), vld1q_f32(ptr + 8), vld1q_f32(ptr + 12)}) {}
+
+  explicit FP32Vec16(float32x4x4_t data) : reg(data) {}
+
+  explicit FP32Vec16(const FP32Vec8 &data) {
+        reg.val[0] = data.reg.val[0]; 
+        reg.val[1] = data.reg.val[1]; 
+        reg.val[2] = data.reg.val[0]; 
+        reg.val[3] = data.reg.val[1]; 
+  }
+
+  explicit FP32Vec16(const FP32Vec16 &data) : reg(data.reg) {}
+
+  explicit FP32Vec16(const FP16Vec8 &v) : FP32Vec16(FP32Vec8(v.reg)) {}
+
+  #ifdef ARM_BF16_SUPPORT
+  explicit FP32Vec16(bfloat16x8x2_t v) : reg({
+    vcvtq_low_f32_bf16(v.val[0]),
+    vcvtq_high_f32_bf16(v.val[0]),
+    vcvtq_low_f32_bf16(v.val[1]),
+    vcvtq_high_f32_bf16(v.val[1])
+  }) {};
+  #endif
+
+  explicit FP32Vec16(const FP32Vec4 &data) {
+    reg.val[0] = data.reg;
+    reg.val[1] = data.reg;
+    reg.val[2] = data.reg;
+    reg.val[3] = data.reg;
+  };
+
+  #ifdef ARM_BF16_SUPPORT
+  explicit FP32Vec16(const BF16Vec16 &v) : reg({
+    vcvtq_low_f32_bf16(v.reg.val[0]),
+    vcvtq_high_f32_bf16(v.reg.val[0]),
+    vcvtq_low_f32_bf16(v.reg.val[1]),
+    vcvtq_high_f32_bf16(v.reg.val[1])
+  }) {};
+
+  explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {};
+  #endif
+
+  explicit FP32Vec16(const FP16Vec16 &v) {
+      reg.val[0] = vcvt_f32_f16(vget_low_f16(v.reg.val[0]));
+      reg.val[1] = vcvt_f32_f16(vget_high_f16(v.reg.val[0]));
+      reg.val[2] = vcvt_f32_f16(vget_low_f16(v.reg.val[1]));
+      reg.val[3] = vcvt_f32_f16(vget_high_f16(v.reg.val[1]));
+  };
+
+  FP32Vec16 operator+(const FP32Vec16 &b) const {
+    return FP32Vec16(float32x4x4_t({
+        vaddq_f32(reg.val[0], b.reg.val[0]),
+        vaddq_f32(reg.val[1], b.reg.val[1]),
+        vaddq_f32(reg.val[2], b.reg.val[2]),
+        vaddq_f32(reg.val[3], b.reg.val[3])}));
+  };
+
+  FP32Vec16 operator*(const FP32Vec16 &b) const {
+    return FP32Vec16(float32x4x4_t({
+        vmulq_f32(reg.val[0], b.reg.val[0]),
+        vmulq_f32(reg.val[1], b.reg.val[1]),
+        vmulq_f32(reg.val[2], b.reg.val[2]),
+        vmulq_f32(reg.val[3], b.reg.val[3])}));
+  };
+
+  FP32Vec16 operator-(const FP32Vec16 &b) const {
+    return FP32Vec16(float32x4x4_t({
+        vsubq_f32(reg.val[0], b.reg.val[0]),
+        vsubq_f32(reg.val[1], b.reg.val[1]),
+        vsubq_f32(reg.val[2], b.reg.val[2]),
+        vsubq_f32(reg.val[3], b.reg.val[3])
+    }));
+  };
+
+  FP32Vec16 operator/(const FP32Vec16 &b) const {
+    return FP32Vec16(float32x4x4_t({
+        vdivq_f32(reg.val[0], b.reg.val[0]),
+        vdivq_f32(reg.val[1], b.reg.val[1]),
+        vdivq_f32(reg.val[2], b.reg.val[2]),
+        vdivq_f32(reg.val[3], b.reg.val[3])
+    }));
+  };
+
+  float reduce_sum() const {
+    AliasReg ar;
+    ar.reg = reg;
+    float answer = 0;
+    unroll_loop<int, VEC_ELEM_NUM>([&answer, &ar](int i) { answer += ar.values[i]; });
+
+    return answer;
+  };
+
+  template <int group_size> float reduce_sub_sum(int idx) {
+    static_assert(VEC_ELEM_NUM % group_size == 0);
+
+    AliasReg ar;
+    ar.reg = reg;
+    float answer = 0;
+    const int start = idx * group_size;
+    unroll_loop<int, group_size>(
+        [&answer, &start, ar](int i) { answer += ar.values[start + i]; });
+
+    return answer;
+  };
+
+  void save(float *ptr) const {
+    vst1q_f32(ptr, reg.val[0]);
+    vst1q_f32(ptr + 4, reg.val[1]);
+    vst1q_f32(ptr + 8, reg.val[2]);
+    vst1q_f32(ptr + 12, reg.val[3]);
+  };
+};
+
+template <typename T> struct VecType { using vec_type = void; };
+
+template <typename T> using vec_t = typename VecType<T>::vec_type;
+
+template <> struct VecType<float> { using vec_type = FP32Vec8; };
+
+template <> struct VecType<c10::Half> { using vec_type = FP16Vec8; };
+
+#ifdef ARM_BF16_SUPPORT
+template <> struct VecType<c10::BFloat16> { using vec_type = BF16Vec8; };
+#endif
+
+template <typename T> void storeFP32(float v, T *ptr) { *ptr = v; }
+
+template <> inline void storeFP32<c10::Half>(float v, c10::Half *ptr) {
+  *reinterpret_cast<__fp16 *>(ptr) = v;
+}
+
+inline FP16Vec16::FP16Vec16(const FP32Vec16 &v) {
+    float16x4_t low_0 = vcvt_f16_f32(v.reg.val[0]);
+    float16x4_t high_0 = vcvt_f16_f32(v.reg.val[1]);
+    float16x4_t low_1 = vcvt_f16_f32(v.reg.val[2]);
+    float16x4_t high_1 = vcvt_f16_f32(v.reg.val[3]);
+
+    reg.val[0] = vcombine_f16(low_0, high_0);
+    reg.val[1] = vcombine_f16(low_1, high_1);
+};
+
+inline FP16Vec8 :: FP16Vec8(const FP32Vec8 &v) {
+    float16x4_t lower_half = vcvt_f16_f32(v.reg.val[0]);
+    float16x4_t upper_half = vcvt_f16_f32(v.reg.val[1]);
+
+    reg = vcombine_f16(lower_half, upper_half);
+};
+
+inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) {
+
+  acc.reg.val[0] = vfmaq_f32(acc.reg.val[0], a.reg.val[0], b.reg.val[0]);
+  acc.reg.val[1] = vfmaq_f32(acc.reg.val[1], a.reg.val[1], b.reg.val[1]);
+  acc.reg.val[2] = vfmaq_f32(acc.reg.val[2], a.reg.val[2], b.reg.val[2]);
+  acc.reg.val[3] = vfmaq_f32(acc.reg.val[3], a.reg.val[3], b.reg.val[3]);
+};
+
+#ifdef ARM_BF16_SUPPORT
+inline void fma(FP32Vec16 &acc, BF16Vec32 &a, BF16Vec32 &b) {
+
+  float32x4_t a0_low = vcvt_f32_bf16(vget_low_bf16(a.reg.val[0]));
+  float32x4_t a0_high = vcvt_f32_bf16(vget_high_bf16(a.reg.val[0]));
+  float32x4_t a1_low = vcvt_f32_bf16(vget_low_bf16(a.reg.val[1]));
+  float32x4_t a1_high = vcvt_f32_bf16(vget_high_bf16(a.reg.val[1]));
+
+  float32x4_t b0_low = vcvt_f32_bf16(vget_low_bf16(b.reg.val[0]));
+  float32x4_t b0_high = vcvt_f32_bf16(vget_high_bf16(b.reg.val[0]));
+  float32x4_t b1_low = vcvt_f32_bf16(vget_low_bf16(b.reg.val[1]));
+  float32x4_t b1_high = vcvt_f32_bf16(vget_high_bf16(b.reg.val[1]));
+
+  acc.reg.val[0] = vfmaq_f32(acc.reg.val[0], a0_low, b0_low);
+  acc.reg.val[1] = vfmaq_f32(acc.reg.val[1], a0_high, b0_high);
+  acc.reg.val[2] = vfmaq_f32(acc.reg.val[2], a1_low, b1_low);
+  acc.reg.val[3] = vfmaq_f32(acc.reg.val[3], a1_high, b1_high);
+};
+#endif
+
+#ifdef ARM_BF16_SUPPORT
+inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) : reg(vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[0]), v.reg.val[1])) {};
+
+inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) : reg({
+    vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[0]), v.reg.val[1]),
+    vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[2]), v.reg.val[3])
+  }){};
+#endif
+
+inline void prefetch(const void *addr) {
+    __builtin_prefetch(addr, 0, 1);
+};
+
+#ifdef ARM_BF16_SUPPORT
+template <>
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) { 
+  *reinterpret_cast<__bf16 *>(ptr) = vcvth_bf16_f32(v);
+};
+#endif
+};
\ No newline at end of file
diff --git a/docs/source/getting_started/arm-installation.rst b/docs/source/getting_started/arm-installation.rst
new file mode 100644
index 0000000000000..7b457df92c11d
--- /dev/null
+++ b/docs/source/getting_started/arm-installation.rst
@@ -0,0 +1,50 @@
+.. _installation_arm:
+
+Installation for ARM CPUs
+=========================
+
+vLLM has been adapted to work on ARM64 CPUs with NEON support, leveraging the CPU backend initially developed for the x86 platform. This guide provides installation instructions specific to ARM. For additional details on supported features, refer to the x86 platform documentation covering:
+
+* CPU backend inference capabilities
+* Relevant runtime environment variables
+* Performance optimization tips
+
+ARM CPU backend currently supports Float32, FP16 and BFloat16 datatypes.
+Contents:
+
+1. :ref:`Requirements <arm_backend_requirements>`
+2. :ref:`Quick Start with Dockerfile <arm_backend_quick_start_dockerfile>`
+3. :ref:`Building from Source <build_arm_backend_from_source>`
+
+.. _arm_backend_requirements:
+
+Requirements
+------------
+
+* **Operating System**: Linux or macOS
+* **Compiler**: gcc/g++ >= 12.3.0 (optional, but recommended)
+* **Instruction Set Architecture (ISA)**: NEON support is required
+
+.. _arm_backend_quick_start_dockerfile:
+
+Quick Start with Dockerfile
+---------------------------
+
+You can quickly set up vLLM on ARM using Docker:
+
+.. code-block:: console
+
+    $ docker build -f Dockerfile.arm -t vllm-cpu-env --shm-size=4g .
+    $ docker run -it \
+                 --rm \
+                 --network=host \
+                 --cpuset-cpus=<cpu-id-list, optional> \
+                 --cpuset-mems=<memory-node, optional> \
+                 vllm-cpu-env
+
+.. _build_arm_backend_from_source:
+
+Building from Source
+--------------------
+
+To build vLLM from source on Ubuntu 22.04 or other Linux distributions, follow a similar process as with x86. Testing has been conducted on AWS Graviton3 instances for compatibility.
diff --git a/docs/source/index.rst b/docs/source/index.rst
index c2afd806c50f9..0692e949f1c77 100644
--- a/docs/source/index.rst
+++ b/docs/source/index.rst
@@ -67,6 +67,7 @@ Documentation
    getting_started/openvino-installation
    getting_started/cpu-installation
    getting_started/gaudi-installation
+   getting_started/arm-installation
    getting_started/neuron-installation
    getting_started/tpu-installation
    getting_started/xpu-installation
diff --git a/examples/offline_inference.py b/examples/offline_inference.py
index 9b758fa2479f6..23cc6e8539431 100644
--- a/examples/offline_inference.py
+++ b/examples/offline_inference.py
@@ -19,4 +19,4 @@
 for output in outputs:
     prompt = output.prompt
     generated_text = output.outputs[0].text
-    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
\ No newline at end of file
diff --git a/requirements-cpu.txt b/requirements-cpu.txt
index 749b03a0603d8..db8ad9d3a015d 100644
--- a/requirements-cpu.txt
+++ b/requirements-cpu.txt
@@ -1,6 +1,7 @@
 # Common dependencies
 -r requirements-common.txt
 
-# Dependencies for x86_64 CPUs
-torch == 2.5.1+cpu; platform_machine != "ppc64le"
-torchvision; platform_machine != "ppc64le"   # required for the image processor of phi3v, this must be updated alongside torch
+# Dependencies for CPUs
+torch==2.5.1+cpu; platform_machine != "ppc64le" and platform_machine != "aarch64" 
+torch==2.5.1; platform_machine == "aarch64"
+torchvision; platform_machine != "ppc64le"  # required for the image processor of phi3v, this must be updated alongside torch
\ No newline at end of file