init upload

cache_autogptq_cuda_256.cpp

@@ -0,0 +1,199 @@
+#include <torch/all.h>
+#include <torch/python.h>
+#include <c10/cuda/CUDAGuard.h>
+
+// adapted from https://github.com/PanQiWei/AutoGPTQ/blob/main/autogptq_extension/cuda_256/autogptq_cuda_256.cpp
+void vecquant8matmul_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros,
+  torch::Tensor g_idx
+);
+
+void vecquant8matmul(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros,
+  torch::Tensor g_idx
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant8matmul_cuda(vec, mat, mul, scales, zeros, g_idx);
+}
+
+void vecquant8matmul_batched_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant8matmul_batched(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant8matmul_batched_cuda(vec, mat, mul, scales, zeros);
+}
+
+void vecquant8matmul_batched_column_compression_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant8matmul_batched_column_compression(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant8matmul_batched_column_compression_cuda(vec, mat, mul, scales, zeros);
+}
+
+void vecquant4matmul_batched_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant4matmul_batched(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant4matmul_batched_cuda(vec, mat, mul, scales, zeros);
+}
+
+void vecquant4matmul_batched_column_compression_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant4matmul_batched_column_compression(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant4matmul_batched_column_compression_cuda(vec, mat, mul, scales, zeros);
+}
+
+void vecquant8matmul_batched_old_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant8matmul_batched_old(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant8matmul_batched_old_cuda(vec, mat, mul, scales, zeros);
+}
+
+
+void vecquant4matmul_batched_old_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant4matmul_batched_old(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant4matmul_batched_old_cuda(vec, mat, mul, scales, zeros);
+}
+
+void vecquant8matmul_batched_column_compression_old_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant8matmul_batched_column_compression_old(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant8matmul_batched_column_compression_old_cuda(vec, mat, mul, scales, zeros);
+}
+
+void vecquant4matmul_batched_column_compression_old_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant4matmul_batched_column_compression_old(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant4matmul_batched_column_compression_old_cuda(vec, mat, mul, scales, zeros);
+}
+
+
+
+void vecquant8matmul_batched_faster_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant8matmul_batched_faster(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant8matmul_batched_faster_cuda(vec, mat, mul, scales, zeros);
+}
+
+
+void vecquant8matmul_batched_faster_old_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant8matmul_batched_faster_old(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant8matmul_batched_faster_old_cuda(vec, mat, mul, scales, zeros);
+}
+
+void vecquant8matmul_batched_column_compression_faster_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant8matmul_batched_column_compression_faster(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant8matmul_batched_column_compression_faster_cuda(vec, mat, mul, scales, zeros);
+}
+
+
+void vecquant8matmul_batched_column_compression_faster_old_cuda(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+);
+
+void vecquant8matmul_batched_column_compression_faster_old(
+  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+  torch::Tensor scales, torch::Tensor zeros
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
+  vecquant8matmul_batched_column_compression_faster_old_cuda(vec, mat, mul, scales, zeros);
+}
+
+
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("vecquant8matmul", &vecquant8matmul, "Vector 8-bit Quantized Matrix Multiplication (CUDA) (desc_act)");
+  m.def("vecquant8matmul_batched", &vecquant8matmul_batched, "Vector 8-bit Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
+  m.def("vecquant8matmul_batched_old", &vecquant8matmul_batched_old, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
+  m.def("vecquant8matmul_batched_faster", &vecquant8matmul_batched_faster, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
+  m.def("vecquant8matmul_batched_faster_old", &vecquant8matmul_batched_faster_old, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
+  m.def("vecquant4matmul_batched_old", &vecquant4matmul_batched_old, "Vector 4-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
+  m.def("vecquant8matmul_batched_column_compression", &vecquant8matmul_batched_column_compression, "Vector 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
+  m.def("vecquant8matmul_batched_column_compression_old", &vecquant8matmul_batched_column_compression_old, "Vector old 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
+  m.def("vecquant8matmul_batched_column_compression_faster", &vecquant8matmul_batched_column_compression_faster, "Vector old 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
+  m.def("vecquant8matmul_batched_column_compression_faster_old", &vecquant8matmul_batched_column_compression_faster_old, "Vector old 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
+  m.def("vecquant4matmul_batched_column_compression_old", &vecquant4matmul_batched_column_compression_old, "Vector old 4-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
+  m.def("vecquant4matmul_batched", &vecquant4matmul_batched, "Vector 4-bit Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
+  m.def("vecquant4matmul_batched_column_compression", &vecquant4matmul_batched_column_compression, "Vector 4-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
+}
+

cache_autogptq_cuda_kernel_256.cu

@@ -0,0 +1,1709 @@
+#define _CRT_SECURE_NO_WARNINGS
+#include <torch/all.h>
+#include <torch/python.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+#include <stdint.h>
+
+#if (defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 700) || defined(USE_ROCM)
+// adapted from https://github.com/PanQiWei/AutoGPTQ/blob/main/autogptq_extension/cuda_256/autogptq_cuda_kernel_256.cu
+__device__ __forceinline__ void atomicAdd(c10::Half* address, c10::Half val) {
+    unsigned int *address_as_ui = reinterpret_cast<unsigned int *>(reinterpret_cast<char *>(address) - (reinterpret_cast<size_t>(address) & 2));
+    unsigned int old = *address_as_ui;
+    unsigned int assumed;
+
+    do {
+        assumed = old;
+        unsigned short hsum = reinterpret_cast<size_t>(address) & 2 ? (old >> 16) : (old & 0xffff);
+        hsum += val;
+        old = reinterpret_cast<size_t>(address) & 2
+                 ? (old & 0xffff) | (hsum << 16)
+                 : (old & 0xffff0000) | hsum;
+        old = atomicCAS(address_as_ui, assumed, old);
+
+    // Note: uses integer comparison to avoid hang in case of NaN (since NaN != NaN)
+    } while (assumed != old);
+}
+__device__ __forceinline__ void atomicAdd(__half* address, c10::Half val) {
+    unsigned int * address_as_ui = (unsigned int *) ((char *)address - ((size_t)address & 2));
+    unsigned int old = *address_as_ui;
+    unsigned int assumed;
+
+    do {
+        assumed = old;
+        __half_raw hsum;
+        hsum.x = (size_t)address & 2 ? (old >> 16) : (old & 0xffff);
+        half tmpres = __hadd(hsum, val);
+        hsum = __half_raw(tmpres);
+        old = (size_t)address & 2 ? (old & 0xffff) | (hsum.x << 16) : (old & 0xffff0000) | hsum.x;
+        old = atomicCAS(address_as_ui, assumed, old);
+    } while (assumed != old);
+}
+#endif
+
+template <typename scalar_t>
+__global__ void VecQuant8MatMulKernel(
+    const  scalar_t* __restrict__ vec,
+    const       int* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const       int* __restrict__ zeros,
+    const       int* __restrict__ g_idx,
+    int batch,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+);
+
+template <typename scalar_t>
+__global__ void VecQuant8BatchMatMulColumnCompressionKernel(
+    const  scalar_t* __restrict__ vec,
+    const       int* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const       int* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+);
+
+template <typename scalar_t>
+__global__ void VecQuant4BatchMatMulColumnCompressionKernel(
+    const  scalar_t* __restrict__ vec,
+    const       int* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const       int* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+);
+
+template <typename scalar_t>
+__global__ void VecQuant8BatchMatMulKernel(
+    const  scalar_t* __restrict__ vec,
+    const       int* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const       int* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+);
+
+template <typename scalar_t>
+__global__ void VecQuant4BatchMatMulKernel(
+    const  scalar_t* __restrict__ vec,
+    const       int* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const       int* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+);
+
+
+
+template <typename scalar_t>
+__global__ void VecQuant8BatchMatMulKernel_old(
+    const  scalar_t* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const  scalar_t* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+);
+
+__global__ void VecQuant8BatchMatMulKernel_faster(
+    const  half* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           half* __restrict__ mul,
+    const  half* __restrict__ scales,
+    const  half* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+);
+
+
+
+__global__ void VecQuant8BatchMatMulKernel_faster_old(
+    const  half* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           half* __restrict__ mul,
+    const  half* __restrict__ scales,
+    const  half* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width
+);
+
+
+template <typename scalar_t>
+__global__ void VecQuant4BatchMatMulKernel_old(
+    const  scalar_t* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const  scalar_t* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+);
+
+
+template <typename scalar_t>
+__global__ void VecQuant8BatchMatMulColumnCompressionKernel_old(
+    const  scalar_t* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const  scalar_t* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+);
+
+__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
+    const  half* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           half* __restrict__ mul,
+    const  half* __restrict__ scales,
+    const  half* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+);
+
+__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster_old(
+    const  half* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           half* __restrict__ mul,
+    const  half* __restrict__ scales,
+    const  half* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+);
+
+
+template <typename scalar_t>
+__global__ void VecQuant4BatchMatMulColumnCompressionKernel_old(
+    const  scalar_t* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const  scalar_t* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+);
+
+
+__global__ void VecQuant8BatchMatMulKernel_faster(
+    const  half* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           half* __restrict__ mul,
+    const  half* __restrict__ scales,
+    const  half* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width
+);
+
+
+__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
+    const  half* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           half* __restrict__ mul,
+    const  half* __restrict__ scales,
+    const  half* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+);
+
+const int BLOCKWIDTH  = 128;
+const int BLOCKHEIGHT8 =  32;
+const int BLOCKHEIGHT4 =  16;
+const int BLOCKHEIGHT_OLD4 =  128;
+//const int BLOCKHEIGHT_OLD8 =  128;
+
+__device__ inline unsigned int as_unsigned(int i) {
+  return *reinterpret_cast<unsigned int*>(&i);
+}
+
+__device__ inline int as_int(int i) {
+  return *reinterpret_cast<int*>(&i);
+}
+
+void vecquant8matmul_batched_column_compression_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int height = vec.size(3);
+  int width = mat.size(3) * 4;
+
+  dim3 blocks(
+    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  AT_DISPATCH_FLOATING_TYPES(
+    vec.type(), "vecquant8matmul_batched_cuda", ([&] {
+      VecQuant8BatchMatMulColumnCompressionKernel<<<blocks, threads>>>(
+        vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
+        scales.data<scalar_t>(), zeros.data<int>(),
+        batch, heads, vec_row, height, width
+      );
+    })
+  );
+
+}
+
+template <typename scalar_t>
+__global__ void VecQuant8BatchMatMulColumnCompressionKernel(
+    const  scalar_t* __restrict__ vec,
+    const       int* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const       int* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+) {
+  int weight_total = batch * heads * height * width / 4;
+  int input_total = batch * heads * vec_row * height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  // h is index of height with step being BLOCKWIDTH
+  int h = BLOCKWIDTH * blockIdx.x;
+  // w is index of width with step being 1
+  int w = BLOCKWIDTH * blockIdx.y + tid;
+  if (w >= width && tid >= height) {
+    return;
+  }
+
+  __shared__ scalar_t blockvec[BLOCKWIDTH];
+  int k;
+  scalar_t w_tmp;
+
+  float weight[BLOCKWIDTH];
+
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
+        int i_w = (w / 4);
+        int w_bit = (w % 4) * 8;
+
+        int w_index = (batch_shift * height + h + k) * width / 4 + i_w;
+        if (w_index >= weight_total || w >= width) {
+          weight[k] = 0;
+        } else {
+          scalar_t scale = scales[batch_shift * height + h + k];
+          scalar_t zero = zeros[batch_shift * height + h + k];
+          w_tmp = ((as_unsigned(mat[w_index]) >> w_bit) & 0xFF);
+          weight[k] = scale * (w_tmp - zero);
+        }
+      }
+
+      scalar_t res;
+      for (int vr = 0; vr < vec_row; ++vr){
+          res = 0;
+        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
+        if (vec_index < input_total) {
+            blockvec[tid] = vec[vec_index];
+        } else {
+            blockvec[tid] = 0;
+        }
+
+        __syncthreads();
+          for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
+          // res is the dot product of BLOCKWIDTH elements (part of width)
+            res += weight[k] * blockvec[k];
+        }
+        // add res to the final result, final matrix shape: (batch, vec_row, width)
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (out_index < out_total) {
+            atomicAdd(&mul[out_index], res);
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+void vecquant8matmul_batched_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int vec_height = vec.size(3);
+  int height = mat.size(2);
+  int width = mat.size(3);
+  int zero_width = zeros.size(2);
+
+  dim3 blocks(
+    (height + BLOCKHEIGHT8 - 1) / BLOCKHEIGHT8,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  AT_DISPATCH_FLOATING_TYPES(
+    vec.type(), "vecquant8matmul_batched_cuda", ([&] {
+      VecQuant8BatchMatMulKernel<<<blocks, threads>>>(
+        vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
+        scales.data<scalar_t>(), zeros.data<int>(),
+        batch, heads, vec_row, vec_height, height, width, zero_width
+      );
+    })
+  );
+
+}
+
+template <typename scalar_t>
+__global__ void VecQuant8BatchMatMulKernel(
+    const  scalar_t* __restrict__ vec,
+    const       int* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const       int* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+) {
+  int weight_total = batch * heads * height * width;
+  int input_total = batch * heads * vec_row * vec_height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  // h is index of height with step being BLOCKHEIGHT8
+  int h = BLOCKHEIGHT8 * blockIdx.x;
+  // w is index of width with step being 1
+  int w = BLOCKWIDTH * blockIdx.y + tid;
+  if (w >= width && tid >= vec_height) {
+    return;
+  }
+
+  __shared__ scalar_t blockvec[BLOCKWIDTH];
+  // i is index of mat of block first row
+  int i = width * h + w;
+  // if (i >= width * height) {
+  //   return;
+  // }
+  int k;
+  scalar_t w_tmp;
+
+  int z_w = w / 4;
+  int z_mod = (w % 4) * 8;
+
+  float weight[BLOCKWIDTH];
+
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      for (k = 0; k <  BLOCKWIDTH && h * 4 + k < vec_height; ++k){
+        int k_w = (k / 4);
+        int k_bit = (k % 4) * 8;
+
+        int w_index = batch_shift * height * width + i + (k_w * width);
+        if (w_index >= weight_total || w >= width) {
+          weight[k] = 0;
+        } else {
+          scalar_t scale = scales[batch_shift * width + w];
+          scalar_t zero;
+          if (zero_width == width) {
+            zero = zeros[batch_shift * width + w];
+          } else {
+            zero = scalar_t(((as_unsigned(zeros[batch_shift * zero_width + z_w]) >> z_mod) & 0xFF) + 1);
+          }
+          w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xFF);
+          weight[k] = scale * (w_tmp - zero);
+        }
+      }
+
+      scalar_t res;
+      for (int vr = 0; vr < vec_row; ++vr){
+          res = 0;
+        int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
+        if (vec_index < input_total) {
+            blockvec[tid] = vec[vec_index];
+        } else {
+            blockvec[tid] = 0;
+        }
+
+        __syncthreads();
+          for (k = 0; k <  BLOCKWIDTH && h * 4 + k < vec_height; ++k){
+          // res is the dot product of BLOCKWIDTH elements (part of width)
+            res += weight[k] * blockvec[k];
+        }
+        // add res to the final result, final matrix shape: (batch, vec_row, width)
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (out_index < out_total) {
+            atomicAdd(&mul[out_index], res);
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+
+void vecquant8matmul_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros,
+  torch::Tensor g_idx
+) {
+  int batch = vec.size(0);
+  int vec_height = vec.size(1);
+  int height = mat.size(0);
+  int width = mat.size(1);
+  int zero_width = zeros.size(1);
+
+  dim3 blocks(
+    (height + BLOCKHEIGHT8 - 1) / BLOCKHEIGHT8,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  AT_DISPATCH_FLOATING_TYPES(
+    vec.type(), "vecquant8matmul_cuda", ([&] {
+      VecQuant8MatMulKernel<<<blocks, threads>>>(
+        vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
+        scales.data<scalar_t>(), zeros.data<int>(), g_idx.data<int>(),
+        batch, vec_height, height, width, zero_width
+      );
+    })
+  );
+}
+
+template <typename scalar_t>
+__global__ void VecQuant8MatMulKernel(
+    const  scalar_t* __restrict__ vec,
+    const       int* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const       int* __restrict__ zeros,
+    const       int* __restrict__ g_idx,
+    int batch,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+) {
+  int h = BLOCKHEIGHT8 * blockIdx.x;
+  int w = BLOCKWIDTH * blockIdx.y + threadIdx.x;
+
+  __shared__ scalar_t blockvec[BLOCKWIDTH];
+  int i = width * h + w;
+  int g_h = h * 4;
+  int k;
+  unsigned int g;
+  scalar_t w_tmp;
+
+  int z_w = w / 4;
+  int z_mod = (w % 4) * 8;
+
+  float weight[BLOCKWIDTH];
+
+  for (k = 0; k <  BLOCKWIDTH; ++k){
+    int k_w = (k / 4);
+    int k_bit = (k % 4) * 8;
+
+      g = as_int(g_idx[g_h + k]);
+      scalar_t scale = scales[g * width + w];
+      scalar_t zero = scalar_t(((as_unsigned(zeros[g * zero_width + z_w]) >> z_mod) & 0xFF) + 1);
+
+      w_tmp = ((as_unsigned(mat[i + (k_w * width)]) >> k_bit) & 0xFF);
+
+    weight[k] = scale * (w_tmp - zero);
+  }
+
+
+  scalar_t res;
+  for (int b = 0; b < batch; ++b){
+      res = 0;
+    blockvec[threadIdx.x] = vec[b * vec_height + blockIdx.x * BLOCKWIDTH + threadIdx.x];
+    __syncthreads();
+    for (k = 0; k <  BLOCKWIDTH; ++k){
+      res += weight[k] * blockvec[k];
+    }
+    atomicAdd(&mul[b * width + w], res);
+    __syncthreads();
+  }
+}
+
+
+
+void vecquant4matmul_batched_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int vec_height = vec.size(3);
+  int height = mat.size(2);
+  int width = mat.size(3);
+  int zero_width = zeros.size(2);
+
+  dim3 blocks(
+    (height + BLOCKHEIGHT4 - 1) / BLOCKHEIGHT4,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  AT_DISPATCH_FLOATING_TYPES(
+    vec.type(), "vecquant4matmul_batched_cuda", ([&] {
+      VecQuant4BatchMatMulKernel<<<blocks, threads>>>(
+        vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
+        scales.data<scalar_t>(), zeros.data<int>(),
+        batch, heads, vec_row, vec_height, height, width, zero_width
+      );
+    })
+  );
+
+}
+
+template <typename scalar_t>
+__global__ void VecQuant4BatchMatMulKernel(
+    const  scalar_t* __restrict__ vec,
+    const       int* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const       int* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+) {
+  int weight_total = batch * heads * height * width;
+  int input_total = batch * heads * vec_row * vec_height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  // h is index of height with step being BLOCKHEIGHT4
+  int h = BLOCKHEIGHT4 * blockIdx.x;
+  // w is index of width with step being 1
+  int w = BLOCKWIDTH * blockIdx.y + tid;
+  if (w >= width && tid >= vec_height) {
+    return;
+  }
+
+  __shared__ scalar_t blockvec[BLOCKWIDTH];
+  // i is index of mat of block first row
+  int i = width * h + w;
+  int k;
+  scalar_t w_tmp;
+
+  int z_w = w / 8;
+  int z_mod = (w % 8) * 4;
+
+  float weight[BLOCKWIDTH];
+
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      for (k = 0; k <  BLOCKWIDTH && h * 8 + k < vec_height; ++k){
+        int k_w = (k / 8);
+        int k_bit = (k % 8) * 4;
+
+        int w_index = batch_shift * height * width + i + (k_w * width);
+        if (w_index >= weight_total || w >= width) {
+          weight[k] = 0;
+        } else {
+          scalar_t scale = scales[batch_shift * width + w];
+          scalar_t zero;
+          if (zero_width == width) {
+            zero = zeros[batch_shift * width + w];
+          } else {
+            zero = scalar_t(((as_unsigned(zeros[batch_shift * zero_width + z_w]) >> z_mod) & 0xF));
+          }
+          w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
+          weight[k] = scale * (w_tmp - zero);
+        }
+      }
+
+      scalar_t res;
+      for (int vr = 0; vr < vec_row; ++vr){
+          res = 0;
+        int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
+        if (vec_index < input_total) {
+            blockvec[tid] = vec[vec_index];
+        } else {
+            blockvec[tid] = 0;
+        }
+
+        __syncthreads();
+          for (k = 0; k <  BLOCKWIDTH && h * 8 + k < vec_height; ++k){
+          // res is the dot product of BLOCKWIDTH elements (part of width)
+            res += weight[k] * blockvec[k];
+        }
+        // add res to the final result, final matrix shape: (batch, vec_row, width)
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (out_index < out_total) {
+            atomicAdd(&mul[out_index], res);
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+
+
+void vecquant4matmul_batched_column_compression_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int height = vec.size(3);
+  int width = mat.size(3) * 8;
+
+  dim3 blocks(
+    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  AT_DISPATCH_FLOATING_TYPES(
+    vec.type(), "vecquant4matmul_batched_cuda", ([&] {
+      VecQuant4BatchMatMulColumnCompressionKernel<<<blocks, threads>>>(
+        vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
+        scales.data<scalar_t>(), zeros.data<int>(),
+        batch, heads, vec_row, height, width
+      );
+    })
+  );
+
+}
+
+template <typename scalar_t>
+__global__ void VecQuant4BatchMatMulColumnCompressionKernel(
+    const  scalar_t* __restrict__ vec,
+    const       int* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const       int* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+) {
+  int weight_total = batch * heads * height * width / 8;
+  int input_total = batch * heads * vec_row * height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  // h is index of height with step being BLOCKWIDTH
+  int h = BLOCKWIDTH * blockIdx.x;
+  // w is index of width with step being 1
+  int w = BLOCKWIDTH * blockIdx.y + tid;
+  if (w >= width && tid >= height) {
+    return;
+  }
+
+  __shared__ scalar_t blockvec[BLOCKWIDTH];
+  int k;
+  scalar_t w_tmp;
+
+  float weight[BLOCKWIDTH];
+
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
+        int i_w = (w / 8);
+        int w_bit = (w % 8) * 4;
+
+        int w_index = (batch_shift * height + h + k) * width / 8 + i_w;
+        if (w_index >= weight_total || w >= width) {
+          weight[k] = 0;
+        } else {
+          scalar_t scale = scales[batch_shift * height + h + k];
+          scalar_t zero = zeros[batch_shift * height + h + k];
+          w_tmp = ((as_unsigned(mat[w_index]) >> w_bit) & 0xF);
+          weight[k] = scale * (w_tmp - zero);
+        }
+      }
+
+      scalar_t res;
+      for (int vr = 0; vr < vec_row; ++vr){
+          res = 0;
+        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
+        if (vec_index < input_total) {
+            blockvec[tid] = vec[vec_index];
+        } else {
+            blockvec[tid] = 0;
+        }
+
+        __syncthreads();
+          for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
+          // res is the dot product of BLOCKWIDTH elements (part of width)
+            res += weight[k] * blockvec[k];
+        }
+        // add res to the final result, final matrix shape: (batch, vec_row, width)
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (out_index < out_total) {
+            atomicAdd(&mul[out_index], res);
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+
+void vecquant8matmul_batched_old_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int vec_height = vec.size(3);
+  int height = mat.size(2);
+  int width = mat.size(3);
+  int zero_width = zeros.size(2);
+
+  dim3 blocks(
+    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  AT_DISPATCH_FLOATING_TYPES(
+    vec.type(), "vecquant8matmul_batched_old_cuda", ([&] {
+      VecQuant8BatchMatMulKernel_old<<<blocks, threads>>>(
+        vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
+        scales.data<scalar_t>(), zeros.data<scalar_t>(),
+        batch, heads, vec_row, vec_height, height, width, zero_width
+      );
+    })
+  );
+}
+
+
+template <typename scalar_t>
+__global__ void VecQuant8BatchMatMulKernel_old(
+    const  scalar_t* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const  scalar_t* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+) {
+  int weight_total = batch * heads * height * width;
+  int input_total = batch * heads * vec_row * vec_height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  // h is index of height with step being BLOCKHEIGHT8
+  int h = BLOCKWIDTH * blockIdx.x;
+  // w is index of width with step being 1
+  int w = BLOCKWIDTH * blockIdx.y + tid;
+  if (w >= width && tid >= vec_height) {
+    return;
+  }
+
+  __shared__ scalar_t blockvec[BLOCKWIDTH];
+  // i is index of mat of block first row
+  int i = width * h + w;
+  int k;
+  scalar_t w_tmp;
+
+  float weight[BLOCKWIDTH];
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      for (k = 0; k <  BLOCKWIDTH && h + k < vec_height; ++k){
+        int k_w = k;
+        int w_index = batch_shift * height * width + i + (k_w * width);
+        if (w_index >= weight_total || w >= width) {
+          weight[k] = 0;
+        } else {
+          scalar_t scale = scales[batch_shift * width + w];
+          scalar_t zero = zeros[batch_shift * width + w];
+          w_tmp = as_unsigned(mat[w_index]);
+          weight[k] = scale * (w_tmp - zero);
+        }
+      }
+
+      scalar_t res;
+      for (int vr = 0; vr < vec_row; ++vr){
+          res = 0;
+        int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
+        if (vec_index < input_total) {
+            blockvec[tid] = vec[vec_index];
+        } else {
+            blockvec[tid] = 0;
+        }
+
+        __syncthreads();
+          for (k = 0; k <  BLOCKWIDTH && h + k < vec_height; ++k){
+          // res is the dot product of BLOCKWIDTH elements (part of width)
+            res += weight[k] * blockvec[k];
+        }
+        // add res to the final result, final matrix shape: (batch, vec_row, width)
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (out_index < out_total) {
+            atomicAdd(&mul[out_index], res);
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+
+
+void vecquant8matmul_batched_faster_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int vec_height = vec.size(3);
+  int height = mat.size(2);
+  int width = mat.size(3);
+  int zero_width = zeros.size(2);
+
+  dim3 blocks(
+    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  VecQuant8BatchMatMulKernel_faster<<<blocks, threads>>>(
+    (half*) vec.data_ptr(),
+    (uint8_t*) mat.data_ptr(),
+    (half*) mul.data_ptr(),
+    (half*) scales.data_ptr(),
+    (half*) zeros.data_ptr(),
+    batch, heads, vec_row, vec_height, height, width, zero_width
+  );
+}
+
+
+
+__global__ void VecQuant8BatchMatMulKernel_faster(
+    const  half* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           half* __restrict__ mul,
+    const  half* __restrict__ scales,
+    const  half* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+) {
+  //int weight_total = batch * heads * height * width;
+  int input_total = batch * heads * vec_row * vec_height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  int h = BLOCKWIDTH * blockIdx.x;
+  int w = BLOCKWIDTH * blockIdx.y + tid;
+  if (w >= width && tid >= height) {
+    return;
+  }
+
+  __shared__ float blockvec[BLOCKWIDTH];
+  int i = width * h + w;
+  int k;
+  float w_tmp;
+
+  float weight[BLOCKWIDTH];
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      for (k = 0; k <  BLOCKWIDTH && h + k < vec_height; ++k){
+        int k_w = k;
+        int w_index = batch_shift * height * width + i + (k_w * width);
+        float scale = __half2float(scales[batch_shift * width + w]);
+        float zero = __half2float(zeros[batch_shift * width + w]);
+        w_tmp = as_unsigned(mat[w_index]);
+        weight[k] = scale *(w_tmp-zero);
+      }
+
+      float res;
+      for (int vr = 0; vr < vec_row; ++vr){
+        res = 0;
+        int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
+        if (vec_index < input_total) {
+            blockvec[tid] = __half2float(vec[vec_index]);
+        } else {
+            blockvec[tid] = 0;
+        }
+        __syncthreads();
+          for (k = 0; k <  BLOCKWIDTH && h + k < vec_height; ++k){
+            float temp_res = weight[k]*blockvec[k];
+            res += temp_res;
+        }
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (out_index < out_total) {
+            atomicAdd(&mul[out_index], __float2half(res));
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+
+
+
+void vecquant8matmul_batched_column_compression_faster_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int height = vec.size(3);
+  int width = mat.size(3);
+
+  dim3 blocks(
+    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  VecQuant8BatchMatMulColumnCompressionKernel_faster<<<blocks, threads>>>(
+    (half*) vec.data_ptr(),
+    (uint8_t*) mat.data_ptr(),
+    (half*) mul.data_ptr(),
+    (half*) scales.data_ptr(),
+    (half*) zeros.data_ptr(),
+    batch, heads, vec_row, height, width
+  );
+
+}
+
+__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
+    const  half* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           half* __restrict__ mul,
+    const  half* __restrict__ scales,
+    const  half* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+) {
+  //int weight_total = batch * heads * height * width;
+  int input_total = batch * heads * vec_row * height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  int h = BLOCKWIDTH * blockIdx.x;
+  int w = BLOCKWIDTH * blockIdx.y + tid;
+  if (w >= width && tid >= height) {
+    return;
+  }
+
+  __shared__ float blockvec[BLOCKWIDTH];
+  int k;
+  float w_tmp;
+  float weight[BLOCKWIDTH];
+
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      for (k = 0; k <  BLOCKWIDTH; ++k){
+        int w_index = (batch_shift * height + h + k) * width  + w;
+        float scale = __half2float(scales[batch_shift * height + h + k]);
+        float zero = __half2float(zeros[batch_shift * height + h + k]);
+        w_tmp = mat[w_index];
+        weight[k] = scale * (w_tmp-zero);
+      }
+
+      float res;
+      for (int vr = 0; vr < vec_row; ++vr){
+        res = 0;
+        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
+        if (vec_index < input_total) {
+            blockvec[tid] = __half2float(vec[vec_index]);
+        } else {
+            blockvec[tid] = 0;
+        }
+        __syncthreads();
+          for (k = 0; k <  BLOCKWIDTH; ++k){
+            res += weight[k]*blockvec[k];
+        }
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (out_index < out_total) {
+            atomicAdd(&mul[out_index], __float2half(res));
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+
+
+void vecquant8matmul_batched_column_compression_old_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int height = vec.size(3);
+  int width = mat.size(3);
+
+  dim3 blocks(
+    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  AT_DISPATCH_FLOATING_TYPES(
+    vec.type(), "vecquant8matmul_batched_column_compression_old_cuda", ([&] {
+      VecQuant8BatchMatMulColumnCompressionKernel_old<<<blocks, threads>>>(
+        vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
+        scales.data<scalar_t>(), zeros.data<scalar_t>(),
+        batch, heads, vec_row, height, width
+      );
+    })
+  );
+
+}
+
+template <typename scalar_t>
+__global__ void VecQuant8BatchMatMulColumnCompressionKernel_old(
+    const  scalar_t* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const  scalar_t* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+) {
+  int weight_total = batch * heads * height * width;
+  int input_total = batch * heads * vec_row * height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  // h is index of height with step being BLOCKWIDTH
+  int h = BLOCKWIDTH * blockIdx.x;
+  // w is index of width with step being 1
+  int w = BLOCKWIDTH * blockIdx.y + tid;
+  if (w >= width && tid >= height) {
+    return;
+  }
+
+  __shared__ scalar_t blockvec[BLOCKWIDTH];
+  int k;
+  scalar_t w_tmp;
+
+  float weight[BLOCKWIDTH];
+
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
+        int w_index = (batch_shift * height + h + k) * width  + w;
+        if (w_index >= weight_total || w >= width) {
+          weight[k] = 0;
+        } else {
+          scalar_t scale = scales[batch_shift * height + h + k];
+          scalar_t zero = zeros[batch_shift * height + h + k];
+          w_tmp = mat[w_index];
+          weight[k] = scale * (w_tmp - zero);
+        }
+      }
+
+      scalar_t res;
+      for (int vr = 0; vr < vec_row; ++vr){
+          res = 0;
+        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
+        if (vec_index < input_total) {
+            blockvec[tid] = vec[vec_index];
+        } else {
+            blockvec[tid] = 0;
+        }
+
+        __syncthreads();
+          for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
+          // res is the dot product of BLOCKWIDTH elements (part of width)
+            res += weight[k] * blockvec[k];
+        }
+        // add res to the final result, final matrix shape: (batch, vec_row, width)
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (out_index < out_total) {
+            atomicAdd(&mul[out_index], res);
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+
+void vecquant4matmul_batched_old_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int vec_height = vec.size(3);
+  int height = mat.size(2);
+  int width = mat.size(3);
+  int zero_width = zeros.size(2);
+
+  dim3 blocks(
+    (height + BLOCKHEIGHT_OLD4 - 1) / BLOCKHEIGHT_OLD4,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  AT_DISPATCH_FLOATING_TYPES(
+    vec.type(), "vecquant4matmul_batched_old_cuda", ([&] {
+      VecQuant4BatchMatMulKernel_old<<<blocks, threads>>>(
+        vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
+        scales.data<scalar_t>(), zeros.data<scalar_t>(),
+        batch, heads, vec_row, vec_height, height, width, zero_width
+      );
+    })
+  );
+
+}
+
+template <typename scalar_t>
+__global__ void VecQuant4BatchMatMulKernel_old(
+    const  scalar_t* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const  scalar_t* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width,
+    int zero_width
+) {
+  int weight_total = batch * heads * height * width;
+  int input_total = batch * heads * vec_row * vec_height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  // h is index of height with step being BLOCKHEIGHT_OLD4
+  int h = BLOCKHEIGHT_OLD4 * blockIdx.x;
+  // w is index of width with step being 1
+  int w = BLOCKWIDTH * blockIdx.y + tid;
+  if (w >= width && tid >= vec_height) {
+    return;
+  }
+
+  __shared__ scalar_t blockvec[BLOCKWIDTH];
+  // i is index of mat of block first row
+  int i = width * h + w;
+  int k;
+  scalar_t w_tmp;
+
+  float weight[BLOCKWIDTH];
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      for (k = 0; k <  BLOCKWIDTH && h*2 + k < vec_height; ++k){
+        int k_w = (k / 2);
+        int k_bit = (k % 2) * 4;
+        int w_index = batch_shift * height * width + i + (k_w * width);
+        if (w_index >= weight_total || w >= width) {
+          weight[k] = 0;
+        } else {
+          scalar_t scale = scales[batch_shift * width + w];
+          scalar_t zero = zeros[batch_shift * width + w];
+          w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
+          weight[k] = scale * (w_tmp - zero);
+        }
+      }
+
+      scalar_t res;
+      for (int vr = 0; vr < vec_row; ++vr){
+          res = 0;
+        int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
+        if (vec_index < input_total) {
+            blockvec[tid] = vec[vec_index];
+        } else {
+            blockvec[tid] = 0;
+        }
+
+        __syncthreads();
+          for (k = 0; k <  BLOCKWIDTH && h*2 + k < vec_height; ++k){
+          // res is the dot product of BLOCKWIDTH elements (part of width)
+            res += weight[k] * blockvec[k];
+        }
+        // add res to the final result, final matrix shape: (batch, vec_row, width)
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (out_index < out_total) {
+            atomicAdd(&mul[out_index], res);
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+
+
+
+
+void vecquant4matmul_batched_column_compression_old_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int height = vec.size(3);
+  int width = mat.size(3);
+
+  dim3 blocks(
+    (height + BLOCKHEIGHT_OLD4 - 1) / BLOCKHEIGHT_OLD4,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  AT_DISPATCH_FLOATING_TYPES(
+    vec.type(), "vecquant4matmul_batched_column_compression_old_cuda", ([&] {
+      VecQuant4BatchMatMulColumnCompressionKernel_old<<<blocks, threads>>>(
+        vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
+        scales.data<scalar_t>(), zeros.data<scalar_t>(),
+        batch, heads, vec_row, height, width
+      );
+    })
+  );
+
+}
+
+template <typename scalar_t>
+__global__ void VecQuant4BatchMatMulColumnCompressionKernel_old(
+    const  scalar_t* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           scalar_t* __restrict__ mul,
+    const  scalar_t* __restrict__ scales,
+    const  scalar_t* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int height,
+    int width
+) {
+  int weight_total = batch * heads * height * width;
+  int input_total = batch * heads * vec_row * height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  // h is index of height with step being BLOCKWIDTH
+  int h = BLOCKHEIGHT_OLD4 * blockIdx.x;
+  // w is index of width with step being 1
+  int w = BLOCKWIDTH * blockIdx.y + tid;
+  if (w >= width && tid >= height) {
+    return;
+  }
+
+  __shared__ scalar_t blockvec[BLOCKWIDTH];
+  int k;
+  scalar_t w_tmp;
+
+  float weight[BLOCKWIDTH];
+
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      for (k = 0; k <  BLOCKWIDTH && h*2 + k < height; ++k){
+        int k_w = (k / 2);
+        int k_bit = (k % 2) * 4;
+        int w_index = (batch_shift * height + h + k) * width  + k_w;
+        if (w_index >= weight_total || w >= width) {
+          weight[k] = 0;
+        } else {
+          scalar_t scale = scales[batch_shift * height + h + k];
+          scalar_t zero = zeros[batch_shift * height + h + k];
+          w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
+          weight[k] = scale * (w_tmp - zero);
+        }
+      }
+
+      scalar_t res;
+      for (int vr = 0; vr < vec_row; ++vr){
+          res = 0;
+        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
+        if (vec_index < input_total) {
+            blockvec[tid] = vec[vec_index];
+        } else {
+            blockvec[tid] = 0;
+        }
+
+        __syncthreads();
+          for (k = 0; k <  BLOCKWIDTH && h*2 + k < height; ++k){
+          // res is the dot product of BLOCKWIDTH elements (part of width)
+            res += weight[k] * blockvec[k];
+        }
+        // add res to the final result, final matrix shape: (batch, vec_row, width)
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (out_index < out_total) {
+            atomicAdd(&mul[out_index], res);
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+
+
+
+
+void vecquant8matmul_batched_faster_old_cuda(
+  torch::Tensor vec,
+  torch::Tensor mat,
+  torch::Tensor mul,
+  torch::Tensor scales,
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2);
+  int vec_height = vec.size(3);
+  int height = mat.size(2);
+  int width = mat.size(3);
+
+  dim3 blocks(
+    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  VecQuant8BatchMatMulKernel_faster_old<<<blocks, threads>>>(
+    (half*) vec.data_ptr(),
+    (uint8_t*) mat.data_ptr(),
+    (half*) mul.data_ptr(),
+    (half*) scales.data_ptr(),
+    (half*) zeros.data_ptr(),
+    batch, heads, vec_row, vec_height, height, width
+  );
+}
+
+
+__global__ void VecQuant8BatchMatMulKernel_faster_old(
+    const  half* __restrict__ vec,
+    const  uint8_t* __restrict__ mat,
+           half* __restrict__ mul,
+    const  half* __restrict__ scales,
+    const  half* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row,
+    int vec_height,
+    int height,
+    int width
+) {
+ int weight_total = batch * heads * height * width;
+  int input_total = batch * heads * vec_row * vec_height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  const int BLOCKWIDTH_half = BLOCKWIDTH/2;
+
+  int h = BLOCKWIDTH * blockIdx.x; //head_dim, dim=-1
+  int w = BLOCKWIDTH * blockIdx.y + tid; //seq-len, +0-256 ,dim=-2
+  /*
+  if (w >= width && tid >= vec_height) {
+    return;
+  }
+  */
+  __shared__ half blockvec[BLOCKWIDTH]; //256
+  int i = width * h + w;
+  int k;
+
+  half w_tmp1 = __float2half(0);
+  half w_tmp2 = __float2half(0);
+
+  half2 weight[BLOCKWIDTH_half];
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      //int zero_index = batch_shift;
+      for (k = 0; k <  BLOCKWIDTH_half; ++k){
+        int w_index1 = batch_shift * height * width + i + (2 * k * width); // [batch,head,h+k, w]
+        int w_index2 = batch_shift * height * width + i + ((2 * k + 1) * width);
+        int zero_index = batch_shift * width + w; // [batch,head, w]
+        if (w_index1 >= weight_total || w >= width || (2 * k + h) >= height) {
+          weight[k] = __float2half2_rn(0);
+        } else {
+            float zero_f=__half2float(zeros[zero_index]);
+            float scale_f= __half2float(scales[zero_index]);
+            if (w_index2 >= weight_total){
+              w_tmp1 = __float2half((as_unsigned(mat[w_index1]) -zero_f)*scale_f);
+              w_tmp2 = __float2half(0);
+              weight[k] = __halves2half2(w_tmp1,w_tmp2);
+              //printf("zero_index is %d w is %d height is %d width is %d w_index1 is %d w_tmp1 is %f w_tmp2 is %f zero is %f scale is %f low is %f high is %f \n ",zero_index,w,height, width,w_index1,__half2float(w_tmp1),__half2float(w_tmp2),zero_f,scale_f,__low2float(weight[k]),__high2float(weight[k]));
+            }else{
+              w_tmp1 = __int2half_rn(as_unsigned(mat[w_index1]));
+              w_tmp2 = __int2half_rn(as_unsigned(mat[w_index2]));
+
+              //weight[k] = __hmul2(__hsub2(__halves2half2(w_tmp1,w_tmp2), __halves2half2(zero,zero)),__halves2half2(scale,scale));
+              weight[k] = __hfma2(__halves2half2(w_tmp1,w_tmp2), __float2half2_rn(scale_f), __float2half2_rn(-(scale_f * zero_f)));
+              //printf("zero_index1 is %d zero_index2 is %d k is %d head is %d w is %d h is %d height is %d width is %d w_index1 is %d w_index2 is %d zero is %f scale is %f low is %f high is %f \n ",zero_index1,zero_index2,k,head,w,h,height, width,w_index1,w_index2,__half2float(zero1),__half2float(scale1),__low2float(weight[k]),__high2float(weight[k]));
+            }
+        }
+      }
+
+
+      for (int vr = 0; vr < vec_row; ++vr){
+        float res=0;
+        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+        if (vec_index < input_total) {
+            //blockvec[tid] = __half2float(vec[vec_index]);// [batch, head, vr, tid(seq_len dim+)]
+            blockvec[tid] = vec[vec_index];
+            //printf("width is %d height is %d h is %d w is %d vec_index is %d out_index is %d vec_row is %d vec_height is %d,vr is %d tid is %d blockvec is %f\n",width,height, h,w,vec_index,out_index,vec_row,vec_height,vr,tid,blockvec[tid]);
+        } else {
+            blockvec[tid] = __float2half(0);
+        }
+        __syncthreads();
+        if (out_index < out_total) {
+          for (k = 0; k <  BLOCKWIDTH_half; ++k){
+            half2 res2 = __hmul2(weight[k],__halves2half2(blockvec[2*k],blockvec[2*k+1]));
+            res += __low2float(res2) + __high2float(res2);
+          }
+          atomicAdd(&mul[out_index], __float2half(res));
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+
+
+void vecquant8matmul_batched_column_compression_faster_old_cuda(
+  torch::Tensor vec,  // [batch,heads, seq_q, seq_v]
+  torch::Tensor mat, // [batch,heads, seq_v, head_dim]
+  torch::Tensor mul,  // [batch,heads, seq_q,head_dim]
+  torch::Tensor scales, // [batch,heads, head_dim]
+  torch::Tensor zeros
+) {
+  int batch = vec.size(0);
+  int heads = vec.size(1);
+  int vec_row = vec.size(2); //ql
+  int height = mat.size(2); //vl
+  int width = mat.size(3); //head_dim
+
+  dim3 blocks(
+    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
+    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
+  );
+  dim3 threads(BLOCKWIDTH);
+
+  VecQuant8BatchMatMulColumnCompressionKernel_faster_old<<<blocks, threads>>>(
+    (half*) vec.data_ptr(),
+    (uint8_t*) mat.data_ptr(),
+    (half*) mul.data_ptr(),
+    (half*) scales.data_ptr(),
+    (half*) zeros.data_ptr(),
+    batch, heads, vec_row, height, width
+  );
+
+}
+
+
+__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster_old(
+    const  half* __restrict__ vec,  // [batch,heads, seq_q, seq_v]
+    const  uint8_t* __restrict__ mat, // [batch,heads, seq_v, head_dim]
+           half* __restrict__ mul, // [batch,heads, seq_q,head_dim]
+    const  half* __restrict__ scales, // [batch,heads, seq_v]
+    const  half* __restrict__ zeros,
+    int batch,
+    int heads,
+    int vec_row, //seq_q
+    int height, //seq_v
+    int width //head_dim
+) {
+  int weight_total = batch * heads * height * width;
+  int input_total = batch * heads * vec_row * height;
+  int out_total = batch * heads * vec_row * width;
+  int tid = threadIdx.x;
+  int h = BLOCKWIDTH * blockIdx.x; // vl
+  int w = BLOCKWIDTH * blockIdx.y + tid; //head_dim + block
+  if (w >= width && tid >= height) {
+    return;
+  }
+  __shared__ half blockvec[BLOCKWIDTH];
+  int k;
+  half w_tmp1 = __float2half(0);
+  half w_tmp2 = __float2half(0);
+  int i = width * h + w;
+  const int BLOCKWIDTH_half = BLOCKWIDTH/2;
+  half2 weight[BLOCKWIDTH_half];
+
+  for (int b = 0; b < batch; ++b){
+    for (int head = 0; head < heads; ++head){
+      int batch_shift = b * heads + head;
+      //int zero_index = batch_shift;
+      for (k = 0; k <  BLOCKWIDTH_half; ++k){
+        int w_index1 = batch_shift * height * width + i + (2 * k) * width; // [batch,head, h+k, w]
+        int w_index2 = batch_shift * height * width + i + ((2 * k + 1) * width);
+        int zero_index1 = batch_shift * height + h + 2*k; // [batch,head, w]
+        int zero_index2 = batch_shift * height + h + 2*k+1; // [batch,head, w]
+
+        if (w_index1 >= weight_total || (2 * k + h)>=height) {
+          weight[k]=__float2half2_rn(0);
+        } else{
+            //int zero_index = batch_shift + h; // [batch,head, w]
+            //float scale_f1 = __half2float(scales[zero_index1]);
+            //float zero_f1 =  __half2float(zeros[zero_index1]);
+            if (w_index2>=weight_total){
+              w_tmp1 = __float2half((as_unsigned(mat[w_index1]) - __half2float(zeros[zero_index1]))* __half2float(scales[zero_index1]));
+              w_tmp2 = __float2half(0);
+              weight[k] = __halves2half2(w_tmp1,w_tmp2);
+              //printf("zero_index is %d k is %d w is %d head is %d height is %d width is %d w_index1 is %d w_tmp1 is %f w_tmp2 is %f zero is %f scale is %f low is %f high is %f \n ",zero_index,k,w,head,height, width,w_index1,__half2float(w_tmp1),__half2float(w_tmp2),zero_f,scale_f,__low2float(weight[k]),__high2float(weight[k]));
+            }else{
+              w_tmp1 = __int2half_rn(as_unsigned(mat[w_index1]));
+              w_tmp2 = __int2half_rn(as_unsigned(mat[w_index2]));
+              half zero1=zeros[zero_index1];
+              half zero2=zeros[zero_index2];
+              half scale1=scales[zero_index1];
+              half scale2=scales[zero_index2];
+              weight[k] = __hmul2(__hsub2(__halves2half2(w_tmp1,w_tmp2), __halves2half2(zero1,zero2)),__halves2half2(scale1,scale2));
+              //weight[k] = __hfma2(__halves2half2(w_tmp1,w_tmp2), __float2half2_rn(scale_f), __float2half2_rn(-(scale_f * zero_f)));
+              //printf("zero_index1 is %d zero_index2 is %d k is %d head is %d w is %d h is %d height is %d width is %d w_index1 is %d w_index2 is %d zero is %f scale is %f low is %f high is %f \n ",zero_index1,zero_index2,k,head,w,h,height, width,w_index1,w_index2,__half2float(zero1),__half2float(scale1),__low2float(weight[k]),__high2float(weight[k]));
+            }
+          }
+       }
+
+
+      for (int vr = 0; vr < vec_row; ++vr){
+        float res=0;
+        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
+        int out_index = (batch_shift * vec_row + vr) * width + w;
+
+        if (vec_index < input_total) {
+            //blockvec[tid] = __half2float(vec[vec_index]);
+            blockvec[tid] = vec[vec_index];
+            //printf("vec_index is %d out_index is %d vec_row is %d ,vr is %d tid is %d blockvec is %f\n",vec_index,out_index,vec_row,vr,tid,blockvec[tid]);
+        } else {
+            blockvec[tid] = __float2half(0);
+            //blockvec[tid] = 0;
+        }
+        __syncthreads();
+        if (out_index < out_total) {
+            for (k = 0; k <  BLOCKWIDTH_half; ++k){
+                half2 res2 = __hmul2(weight[k],__halves2half2(blockvec[2*k],blockvec[2*k+1]));
+                res += __low2float(res2) + __high2float(res2);
+            }
+            atomicAdd(&mul[out_index], __float2half(res));
+        }
+        __syncthreads();
+      }
+    }
+  }
+}
+

config.json

@@ -0,0 +1,38 @@
+{
+  "architectures": [
+    "QWenLMHeadModel"
+  ],
+  "attn_dropout_prob": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_qwen.QWenConfig",
+    "AutoModelForCausalLM": "modeling_qwen.QWenLMHeadModel"
+  },
+  "bf16": false,
+  "emb_dropout_prob": 0.0,
+  "fp16": false,
+  "fp32": false,
+  "hidden_size": 8192,
+  "initializer_range": 0.02,
+  "intermediate_size": 49152,
+  "kv_channels": 128,
+  "layer_norm_epsilon": 1e-06,
+  "max_position_embeddings": 32768,
+  "model_type": "qwen",
+  "no_bias": true,
+  "num_attention_heads": 64,
+  "num_hidden_layers": 80,
+  "onnx_safe": null,
+  "rope_theta": 1000000,
+  "rotary_emb_base": 1000000,
+  "rotary_pct": 1.0,
+  "scale_attn_weights": true,
+  "seq_length": 32768,
+  "tie_word_embeddings": false,
+  "tokenizer_class": "QWenTokenizer",
+  "transformers_version": "4.32.0",
+  "use_cache": true,
+  "use_dynamic_ntk": false,
+  "use_flash_attn": "auto",
+  "use_logn_attn": false,
+  "vocab_size": 152064
+}

configuration_qwen.py

@@ -0,0 +1,72 @@
+# Copyright (c) Alibaba Cloud.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from transformers import PretrainedConfig
+
+
+class QWenConfig(PretrainedConfig):
+    model_type = "qwen"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        emb_dropout_prob=0.0,
+        attn_dropout_prob=0.0,
+        layer_norm_epsilon=1e-6,
+        initializer_range=0.02,
+        max_position_embeddings=8192,
+        scale_attn_weights=True,
+        use_cache=True,
+        bf16=False,
+        fp16=False,
+        fp32=False,
+        kv_channels=128,
+        rotary_pct=1.0,
+        rotary_emb_base=10000,
+        use_dynamic_ntk=True,
+        use_logn_attn=True,
+        use_flash_attn="auto",
+        intermediate_size=22016,
+        no_bias=True,
+        tie_word_embeddings=False,
+        use_cache_quantization=False,
+        use_cache_kernel=False,
+        softmax_in_fp32=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.emb_dropout_prob = emb_dropout_prob
+        self.attn_dropout_prob = attn_dropout_prob
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.scale_attn_weights = scale_attn_weights
+        self.use_cache = use_cache
+        self.max_position_embeddings = max_position_embeddings
+        self.bf16 = bf16
+        self.fp16 = fp16
+        self.fp32 = fp32
+        self.kv_channels = kv_channels
+        self.rotary_pct = rotary_pct
+        self.rotary_emb_base = rotary_emb_base
+        self.use_dynamic_ntk = use_dynamic_ntk
+        self.use_logn_attn = use_logn_attn
+        self.use_flash_attn = use_flash_attn
+        self.no_bias = no_bias
+        self.use_cache_quantization = use_cache_quantization
+        self.use_cache_kernel = use_cache_kernel
+        self.softmax_in_fp32 = softmax_in_fp32
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs
+        )
+

cpp_kernels.py

@@ -0,0 +1,56 @@
+from torch.utils import cpp_extension
+import pathlib
+import os
+import subprocess
+
+def _get_cuda_bare_metal_version(cuda_dir):
+    raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"],
+                                         universal_newlines=True)
+    output = raw_output.split()
+    release_idx = output.index("release") + 1
+    release = output[release_idx].split(".")
+    bare_metal_major = release[0]
+    bare_metal_minor = release[1][0]
+
+    return raw_output, bare_metal_major, bare_metal_minor
+
+def _create_build_dir(buildpath):
+    try:
+        os.mkdir(buildpath)
+    except OSError:
+        if not os.path.isdir(buildpath):
+            print(f"Creation of the build directory {buildpath} failed")
+
+# Check if cuda 11 is installed for compute capability 8.0
+cc_flag = []
+_, bare_metal_major, bare_metal_minor = _get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
+if int(bare_metal_major) >= 11:
+    cc_flag.append('-gencode')
+    cc_flag.append('arch=compute_80,code=sm_80')
+    if int(bare_metal_minor) >= 7:
+        cc_flag.append('-gencode')
+        cc_flag.append('arch=compute_90,code=sm_90')
+
+# Build path
+srcpath = pathlib.Path(__file__).parent.absolute()
+buildpath = srcpath / 'build'
+_create_build_dir(buildpath)
+
+def _cpp_extention_load_helper(name, sources, extra_cuda_flags):
+    return cpp_extension.load(
+        name=name,
+        sources=sources,
+        build_directory=buildpath,
+        extra_cflags=['-O3', ],
+        extra_cuda_cflags=['-O3',
+                           '-gencode', 'arch=compute_70,code=sm_70',
+                           '--use_fast_math'] + extra_cuda_flags + cc_flag,
+        verbose=1
+    )
+
+extra_flags = []
+
+cache_autogptq_cuda_256_sources = ["./cache_autogptq_cuda_256.cpp",
+           "./cache_autogptq_cuda_kernel_256.cu"]
+cache_autogptq_cuda_256 = _cpp_extention_load_helper("cache_autogptq_cuda_256", cache_autogptq_cuda_256_sources, extra_flags)
+

generation_config.json

@@ -0,0 +1,12 @@
+{
+  "chat_format": "chatml",
+  "do_sample": true,
+  "eos_token_id": 151643,
+  "max_new_tokens": 512,
+  "max_window_size": 6144,
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.1,
+  "top_k": 0,
+  "top_p": 0.8,
+  "transformers_version": "4.31.0"
+}

model-00001-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71f121603f83009d633cfc65772d2e9f3d321e1aeb9c0afbf775a82b5f1d2f30
+size 2491416712

model-00002-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fd6534b1e349606cbccd7b9b1810a52b6e32a225a470948d6528ee109d99952
+size 1744929752

model-00003-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f3c0e23f4e852e8d0e6e264aa4536148cb3573e37feb919f6c52e1234040ea4d
+size 1744913264

model-00004-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36e66f847e111e505066107c3ed63ceca62223fef7ab4a74f612ad66c10c2c00
+size 1744913264

model-00005-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca56ba16415e79fc0a749a01ee30aa69c252d5bfa839a3392e76e8f6fdec0bd2
+size 1744913264

model-00006-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1b24bc0671601ee40cd0b1078dad4600ce982eefeac5b0e06e100a511246d24
+size 1744913264

model-00007-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f20792548d3c809a9464c7127feabc8d149463ef379e97317f9a81e1cb657967
+size 1744913264

model-00008-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f37308c137a3d8f9ee859d122f73a094e5082021e170d9c60e43bb4945f26089
+size 1744913264

model-00009-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4cefda2047e72ac4b0b6b0d809ea4e09787a9010fa51bff74c4a3b6ee0545c8
+size 1744913264

model-00010-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b69dec5dff13f0edf242a641fe7793b9a02422e855ee62966b80dd1c645db3ea
+size 1744913264

model-00011-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9174d918a5de071b5dde02b9cbef792ec3a25c4ea7361e46244c13ce9ec38bf9
+size 1744913256

model-00012-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82fc4256621650e7ae22b90a469a7f210937ba408696b081019a2fe1b2078b75
+size 1744913272

model-00013-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebd93c40b73403e918cf6629958775a5e820227fef0dff42ff952be96949d302
+size 1744913272

model-00014-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da248c713a8c19799f31c9b820d447e9c946e385174848725f54bcba6154ef9a
+size 1744913272

model-00015-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70a3e168d90b44810803d911b82eab009fd9002eb0ac70cfa96ad6b0d4bc12fc
+size 1744913272

model-00016-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfd19fd25338ff5d76f2759a033b372c465afa00422d457c1a1c2bd7be585630
+size 1744913272

model-00017-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c74d65b8a57ac605ae334fcc00c593a468ab5557a7dfa4bb0b9d474d6723e158
+size 1744913272

model-00018-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fc676db6a69a10258def9c0c8bf087de7b75ce1392d2ba65e9302a280c56071
+size 1744913272

model-00019-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:756820512afb119583a05e4306960723b1fb7fb580185bc5309cd5c09459ae99
+size 1744913272

model-00020-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c96634dc837eacddb605510e152d1dd2bf89e2774ad35763398fbc7bc019aa3e
+size 1744913272

model-00021-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8237207c1603a404773576c6725a879c5173429d84651f511f1734fb404cb617
+size 1744913272

model-00022-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ba628bf32cf269dbdba1857a0b5df8e49957093e14751c5612950b919fc4033
+size 1744913272

model-00023-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:081141c800066f30f94c9abd16d55c7215de878d031032ad5674b13d2f4702b8
+size 1744913272

model-00024-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3bc87baabf16a6ebe2ccd673b1065090fea3c5a171dc85b35850a5bbc1db5f2
+size 1744913272

model-00025-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd74400e8aee78e08ef850e734f01781cec92efb6ac922d94eabcdddd0eb6757
+size 1744913272

model-00026-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ed18c462ffca5925fa1d8f3022bc98b9387014b08d86bd4fc9e8d1775ab9be1
+size 1744913272

model-00027-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d85f4bd13d153a1ad4c3a0843d47c781cadfaeb48ee405f99f34dab080f1be78
+size 1744913272

model-00028-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ee6250f0ab45207455bfb6f80d63ef24adb939d8f62728db5a69a8039cd9287
+size 1744913272

model-00029-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:800ac6b701068494837aca3d855c65d1150dbf0550e1e2d503878764c43f3ce7
+size 1744913272

model-00030-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ea296cc93a0a7ae3a641dea7fc044ed5120ad5084d05d354c3abd77caa41a8a
+size 1744913272

model-00031-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6910ca8a9f4c4eca3e9fe23c4fd754be805ec0e44fa5e5a721c4131706c3d8c9
+size 1744913272

model-00032-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac4d97568aff8bc4e29f661a8f73a89615376d424f9fa07244b0715aea4ac7f6
+size 1744913272

model-00033-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b8157bfcc79b1c8941cc2a114164dabe63e71904f8dee3fb74d5045db18c75f
+size 1744913272

model-00034-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:527d1f594f49efd41f8c41c4e7d9ecdd6359c85b75495dd20ab6fe91aaaddb16
+size 1744913272

model-00035-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61fb6a53bf799663ecdb10e681a050d28c08ad84f490be4a0178ef1cc12cfa09
+size 1744913272

model-00036-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:178a50a939136346da409b3a1736b1ce6c59c3b8ad5eba073c9b5c1bb288e799
+size 1744913272

model-00037-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ca4e666cb67e4cb4a8b9e7783e3a42931c1e30fb5ac56cbf2bad9d09361fe12
+size 1744913272

model-00038-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82c44e2af5d8c76ba174bd4f7d98c5321d565db863a78f1421c293d374bf7469
+size 1744913272

model-00039-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3496660eb1e7fe89f6e6a811655c3f229b0830590f3144693fa75176ff662da1
+size 1744913272

model-00040-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54a1f81740344cee132830674fc120c3219623b37ed05ffeb188f450582fec88
+size 1744913272

model-00041-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:065403c6e1dc4e6b71cb5513099391021618f78d10083baebefa5eb83697846e
+size 1744913272

model-00042-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a2ff97ec55237be3de68968e9c6230cb651a70db52d636be79d8d1d721b1210
+size 1744913272

model-00043-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1eda37118b8ab66c686b227a2707cf1ae3a9409ee6b1a3fca14538792e359861
+size 1744913272

model-00044-of-00082.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d7ab4f27b2de838596cfee05aaeaa8961e4c288010585720c8e2db06a14126b
+size 1744913272