diff --git a/onnxruntime/core/providers/vsinpu/patches/mlas_crosscompiling.patch b/onnxruntime/core/providers/vsinpu/patches/mlas_crosscompiling.patch
index f55b593fd8..95a4e4650e 100644
--- a/onnxruntime/core/providers/vsinpu/patches/mlas_crosscompiling.patch
+++ b/onnxruntime/core/providers/vsinpu/patches/mlas_crosscompiling.patch
@@ -1,8 +1,8 @@
 diff --git a/cmake/onnxruntime_mlas.cmake b/cmake/onnxruntime_mlas.cmake
-index c02ac2096d..2bc51298f0 100644
+index 10c307b3b9..a52bf71c4d 100644
 --- a/cmake/onnxruntime_mlas.cmake
 +++ b/cmake/onnxruntime_mlas.cmake
-@@ -361,7 +361,7 @@ else()
+@@ -370,7 +370,7 @@ else()
          )
          set_source_files_properties(${MLAS_SRC_DIR}/sqnbitgemm_kernel_neon_int8.cpp
                                      PROPERTIES COMPILE_FLAGS " -march=armv8.2-a+dotprod")
@@ -12,10 +12,10 @@ index c02ac2096d..2bc51298f0 100644
              ${mlas_platform_srcs}
              ${MLAS_SRC_DIR}/aarch64/HalfGemmKernelNeon.S
 diff --git a/onnxruntime/core/mlas/inc/mlas.h b/onnxruntime/core/mlas/inc/mlas.h
-index e46105324a..414c46a1ce 100644
+index 28ae64c4d5..0c77e0ca78 100644
 --- a/onnxruntime/core/mlas/inc/mlas.h
 +++ b/onnxruntime/core/mlas/inc/mlas.h
-@@ -82,6 +82,9 @@ Abstract:
+@@ -83,6 +83,9 @@ Abstract:
 
  #if (!defined(_MSC_VER)) || (_MSC_VER >= 1930)
  #if defined(MLAS_TARGET_ARM64) || defined(MLAS_TARGET_ARM64EC)
@@ -25,7 +25,7 @@ index e46105324a..414c46a1ce 100644
  #if !defined(__APPLE__)
  // Had to temporary disable fp16 under APPLE ARM64, as compiling
  // the source files require a hardware specific compilation flag.
-@@ -90,6 +93,7 @@ Abstract:
+@@ -91,6 +94,7 @@ Abstract:
 
  #define MLAS_F16VEC_INTRINSICS_SUPPORTED
 
@@ -33,7 +33,7 @@ index e46105324a..414c46a1ce 100644
  #endif //
  #endif // ARM64
  #endif // Visual Studio 16 or earlier does not support fp16 intrinsic
-@@ -1635,6 +1639,7 @@ MlasHalfGemmConvertPackB(
+@@ -1644,6 +1648,7 @@ MlasHalfGemmConvertPackB(
      );
 
  #if defined(__aarch64__) && defined(__linux__)
@@ -41,7 +41,7 @@ index e46105324a..414c46a1ce 100644
  /**
   * @brief Whether current CPU supports Bfloat16(bf16) acceleration.
   */
-@@ -1746,6 +1751,7 @@ MlasSBGemmPackBSize(size_t N, size_t K);
+@@ -1755,6 +1760,7 @@ MlasSBGemmPackBSize(size_t N, size_t K);
  void MLASCALL
  MlasSBGemmConvertPackB(size_t N, size_t K, const float* B, size_t ldb, void* PackedB);
  #endif
@@ -50,10 +50,10 @@ index e46105324a..414c46a1ce 100644
  /**
   * @brief Indirect Depthwise convolution for fp16
 diff --git a/onnxruntime/core/mlas/lib/mlasi.h b/onnxruntime/core/mlas/lib/mlasi.h
-index 4239e2ecae..3df7e5573d 100644
+index 0533a5e49b..c18bf7f90d 100644
 --- a/onnxruntime/core/mlas/lib/mlasi.h
 +++ b/onnxruntime/core/mlas/lib/mlasi.h
-@@ -361,6 +361,7 @@ size_t
+@@ -377,6 +377,7 @@ size_t
  #else
 
  #if defined(__aarch64__) && defined(__linux__)
@@ -61,7 +61,7 @@ index 4239e2ecae..3df7e5573d 100644
  typedef size_t(MLASCALL MLAS_SBGEMM_FLOAT_KERNEL)(
      const float* A,
      const bfloat16_t* B,
-@@ -373,6 +374,7 @@ typedef size_t(MLASCALL MLAS_SBGEMM_FLOAT_KERNEL)(
+@@ -389,6 +390,7 @@ typedef size_t(MLASCALL MLAS_SBGEMM_FLOAT_KERNEL)(
      const float* Bias
  );
  #endif
@@ -69,7 +69,7 @@ index 4239e2ecae..3df7e5573d 100644
 
  typedef
  size_t
-@@ -763,8 +765,10 @@ extern "C" {
+@@ -796,8 +798,10 @@ extern "C" {
      MLAS_GEMM_FLOAT_KERNEL MlasSgemmKernelZero;
      MLAS_GEMM_FLOAT_KERNEL MlasSgemmKernelAdd;
  #if defined(__aarch64__) && defined(__linux__)
@@ -80,7 +80,7 @@ index 4239e2ecae..3df7e5573d 100644
  #endif
      MLAS_GEMM_DOUBLE_KERNEL MlasDgemmKernelZero;
      MLAS_GEMM_DOUBLE_KERNEL MlasDgemmKernelAdd;
-@@ -899,8 +903,10 @@ extern "C" {
+@@ -946,8 +950,10 @@ extern "C" {
  #define MLAS_QGEMM_THREAD_COMPLEXITY                65536
 
  #if defined(__aarch64__) && defined(__linux__)
@@ -91,26 +91,12 @@ index 4239e2ecae..3df7e5573d 100644
 
  //
  // Single-threaded single precision matrix/matrix multiply operation.
-@@ -2570,4 +2576,3 @@ MlasPackInt4Elements(uint8_t* Output, UnpackedType ValueLow, UnpackedType ValueH
-     static_assert(std::is_same_v<UnpackedType, uint8_t> || std::is_same_v<UnpackedType, int8_t>);
-     *Output = static_cast<uint8_t>(((ValueHigh & 0xF) << 4) | (ValueLow & 0xF));
- }
--
 diff --git a/onnxruntime/core/mlas/lib/platform.cpp b/onnxruntime/core/mlas/lib/platform.cpp
-index ed437f20f7..8c9d0a75fd 100644
+index b3c9461293..424c3b0441 100644
 --- a/onnxruntime/core/mlas/lib/platform.cpp
 +++ b/onnxruntime/core/mlas/lib/platform.cpp
-@@ -20,7 +20,7 @@ Abstract:
- #include <thread>
- #include <mutex>
-
--#if defined(MLAS_TARGET_POWER)
-+#if defined(MLAS_TARGET_POWER)
- #if defined(__linux__)
- #include <sys/auxv.h>
- #elif defined(_AIX)
-@@ -536,7 +536,7 @@ Return Value:
-         this->QNBitGemmDispatch = &MlasSQNBitGemmDispatchNeon;
+@@ -574,7 +574,7 @@ Return Value:
+         this->ConvSymS8S8Dispatch = &MlasConvSymS8DispatchDot;
      }
 
 -#if defined(__linux__)
@@ -137,10 +123,10 @@ index de7fd72fad..4f75dbd6fa 100644
 +#endif
  #endif  // defined(__aarch64__) && defined(__linux__)
 diff --git a/onnxruntime/core/providers/cpu/math/matmul.cc b/onnxruntime/core/providers/cpu/math/matmul.cc
-index 6a71283f9d..d8bd348854 100644
+index 2c6d23e4de..61aaacdfd6 100644
 --- a/onnxruntime/core/providers/cpu/math/matmul.cc
 +++ b/onnxruntime/core/providers/cpu/math/matmul.cc
-@@ -132,7 +132,7 @@ Status MatMul<T>::Compute(OpKernelContext* ctx) const {
+@@ -133,7 +133,7 @@ Status MatMul<T>::Compute(OpKernelContext* ctx) const {
 
    return Status::OK();
  }
@@ -149,7 +135,7 @@ index 6a71283f9d..d8bd348854 100644
  bool GemmPackBBfloat16(AllocatorPtr& alloc,
                         const Tensor& tensor_b,
                         bool trans_b,
-@@ -180,6 +180,7 @@ Status MatMul<float>::PrePack(const Tensor& tensor, int input_idx, /*out*/ Alloc
+@@ -181,6 +181,7 @@ Status MatMul<float>::PrePack(const Tensor& tensor, int input_idx, /*out*/ Alloc
    if (input_idx == 1) {
      size_t packed_b_size;
  #if defined(__aarch64__) && defined(__linux__)
@@ -157,7 +143,7 @@ index 6a71283f9d..d8bd348854 100644
      size_t dim1 = 0;
      size_t dim2 = 0;
      TensorShape b_shape = tensor.Shape();
-@@ -192,6 +193,7 @@ Status MatMul<float>::PrePack(const Tensor& tensor, int input_idx, /*out*/ Alloc
+@@ -193,6 +194,7 @@ Status MatMul<float>::PrePack(const Tensor& tensor, int input_idx, /*out*/ Alloc
      if (use_fastmath_mode_ && (trans_b_attr_ == 0) && ((dim1 * dim2) >= kFastMathModeKernelsizeThreshold)) {
        is_packed = GemmPackBBfloat16(alloc, tensor, trans_b_attr_ != 0, packed_b_, packed_b_size, b_shape_);
      } else
@@ -165,7 +151,7 @@ index 6a71283f9d..d8bd348854 100644
  #endif
      {
        is_packed = GemmPackBFp32(alloc, tensor, trans_b_attr_ != 0, packed_b_, packed_b_size, b_shape_);
-@@ -257,6 +259,7 @@ Status MatMul<float>::Compute(OpKernelContext* ctx) const {
+@@ -259,6 +261,7 @@ Status MatMul<float>::Compute(OpKernelContext* ctx) const {
    const size_t lda = helper.Lda(trans_a);
    const size_t ldb = helper.Ldb(trans_b);
  #if defined(__aarch64__) && defined(__linux__)
@@ -173,7 +159,7 @@ index 6a71283f9d..d8bd348854 100644
    if (use_fastmath_mode_ && !trans_b && ((N * K) >= kFastMathModeKernelsizeThreshold)) {
      std::vector<MLAS_SBGEMM_DATA_PARAMS> data(max_len);
      for (size_t i = 0; i < max_len; i++) {
-@@ -273,6 +276,7 @@ Status MatMul<float>::Compute(OpKernelContext* ctx) const {
+@@ -275,6 +278,7 @@ Status MatMul<float>::Compute(OpKernelContext* ctx) const {
      }
      MlasSBGemmBatch(M, N, K, max_len, data.data(), thread_pool);
    } else