mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2025-11-15 11:17:31 +00:00
CUDA: noncont MMVQ + batched bs1 MUL_MAT_ID (#13014)
* CUDA: noncont MMVQ + batched bs1 MUL_MAT_ID * fix logic for RoPE support, CUDA graphs
This commit is contained in:
Johannes Gäßler
@@ -1,6 +1,9 @@
|
||||
#include "mmvq.cuh"
|
||||
#include "quantize.cuh"
|
||||
#include "vecdotq.cuh"
|
||||
|
||||
#include <cstdint>
|
||||
|
||||
typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);
|
||||
|
||||
static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
|
||||
@@ -73,9 +76,9 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
|
||||
return MMVQ_PARAMETERS_GENERIC;
|
||||
}
|
||||
|
||||
static constexpr __host__ __device__ int calc_nwarps(int ncols_y, mmvq_parameter_table_id table_id) {
|
||||
static constexpr __host__ __device__ int calc_nwarps(int ncols_dst, mmvq_parameter_table_id table_id) {
|
||||
if (table_id == MMVQ_PARAMETERS_GENERIC) {
|
||||
switch (ncols_y) {
|
||||
switch (ncols_dst) {
|
||||
case 1:
|
||||
case 2:
|
||||
case 3:
|
||||
@@ -90,7 +93,7 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_y, mmvq_paramete
|
||||
return 1;
|
||||
}
|
||||
} else if (table_id == MMVQ_PARAMETERS_GCN) {
|
||||
switch (ncols_y) {
|
||||
switch (ncols_dst) {
|
||||
case 1:
|
||||
case 2:
|
||||
case 3:
|
||||
@@ -107,9 +110,9 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_y, mmvq_paramete
|
||||
return 1;
|
||||
}
|
||||
|
||||
static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int table_id) {
|
||||
static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id) {
|
||||
if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN) {
|
||||
switch (ncols_y) {
|
||||
switch (ncols_dst) {
|
||||
case 1:
|
||||
return 1;
|
||||
case 2:
|
||||
@@ -127,19 +130,21 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int ta
|
||||
return 1;
|
||||
}
|
||||
|
||||
template <ggml_type type, int ncols_y>
|
||||
template <ggml_type type, int ncols_dst>
|
||||
// tell the compiler to use as many registers as it wants, see nwarps definition below
|
||||
__launch_bounds__(calc_nwarps(ncols_y, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
|
||||
__launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
|
||||
static __global__ void mul_mat_vec_q(
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
|
||||
const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, float * __restrict__ dst,
|
||||
const int ncols_x, const int nchannels_y, const int stride_row_x, const int stride_col_y, const int stride_col_dst,
|
||||
const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
|
||||
const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
|
||||
|
||||
constexpr int qk = ggml_cuda_type_traits<type>::qk;
|
||||
constexpr int qi = ggml_cuda_type_traits<type>::qi;
|
||||
constexpr int vdr = get_vdr_mmvq(type);
|
||||
constexpr mmvq_parameter_table_id table_id = get_device_table_id();
|
||||
constexpr int nwarps = calc_nwarps(ncols_y, table_id);
|
||||
constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_y, table_id);
|
||||
constexpr int nwarps = calc_nwarps(ncols_dst, table_id);
|
||||
constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id);
|
||||
constexpr int warp_size = ggml_cuda_get_physical_warp_size();
|
||||
|
||||
constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
|
||||
@@ -147,13 +152,21 @@ static __global__ void mul_mat_vec_q(
|
||||
const int tid = warp_size*threadIdx.y + threadIdx.x;
|
||||
const int row0 = rows_per_cuda_block*blockIdx.x;
|
||||
const int blocks_per_row_x = ncols_x / qk;
|
||||
const int blocks_per_col_y = nrows_y / QK8_1;
|
||||
constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
|
||||
|
||||
// partial sum for each thread
|
||||
float tmp[ncols_y][rows_per_cuda_block] = {{0.0f}};
|
||||
// The MUL_MAT_ID code path with ids != nullptr is only implemetned for ncols_dst == 1.
|
||||
const int channel_dst = blockIdx.y;
|
||||
const int channel_x = ncols_dst == 1 && ids ? ids[channel_dst] : channel_dst / channel_ratio;
|
||||
const int channel_y = ncols_dst == 1 && ids ? channel_dst % nchannels_y : channel_dst;
|
||||
const int sample_dst = blockIdx.z;
|
||||
const int sample_x = sample_dst / sample_ratio;
|
||||
const int sample_y = sample_dst;
|
||||
|
||||
const block_q8_1 * y = (const block_q8_1 *) vy;
|
||||
// partial sum for each thread
|
||||
float tmp[ncols_dst][rows_per_cuda_block] = {{0.0f}};
|
||||
|
||||
const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
|
||||
const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
|
||||
|
||||
for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
|
||||
const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx
|
||||
@@ -162,18 +175,19 @@ static __global__ void mul_mat_vec_q(
|
||||
const int kqs = vdr * (tid % (qi/vdr));
|
||||
|
||||
#pragma unroll
|
||||
for (int j = 0; j < ncols_y; ++j) {
|
||||
for (int j = 0; j < ncols_dst; ++j) {
|
||||
#pragma unroll
|
||||
for (int i = 0; i < rows_per_cuda_block; ++i) {
|
||||
tmp[j][i] += vec_dot_q_cuda(vx, &y[j*blocks_per_col_y + kby], (row0 + i)*blocks_per_row_x + kbx, kqs);
|
||||
tmp[j][i] += vec_dot_q_cuda(
|
||||
vx, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
__shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][warp_size];
|
||||
__shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
|
||||
if (threadIdx.y > 0) {
|
||||
#pragma unroll
|
||||
for (int j = 0; j < ncols_y; ++j) {
|
||||
for (int j = 0; j < ncols_dst; ++j) {
|
||||
#pragma unroll
|
||||
for (int i = 0; i < rows_per_cuda_block; ++i) {
|
||||
tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
|
||||
@@ -185,9 +199,11 @@ static __global__ void mul_mat_vec_q(
|
||||
return;
|
||||
}
|
||||
|
||||
dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;
|
||||
|
||||
// sum up partial sums and write back result
|
||||
#pragma unroll
|
||||
for (int j = 0; j < ncols_y; ++j) {
|
||||
for (int j = 0; j < ncols_dst; ++j) {
|
||||
#pragma unroll
|
||||
for (int i = 0; i < rows_per_cuda_block; ++i) {
|
||||
#pragma unroll
|
||||
@@ -197,88 +213,121 @@ static __global__ void mul_mat_vec_q(
|
||||
tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
|
||||
}
|
||||
|
||||
if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < (unsigned)nrows_dst)) {
|
||||
dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
|
||||
if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + int(threadIdx.x) < stride_col_dst)) {
|
||||
dst[j*stride_col_dst + threadIdx.x] = tmp[j][threadIdx.x];
|
||||
}
|
||||
}
|
||||
|
||||
GGML_UNUSED(nrows_x);
|
||||
}
|
||||
|
||||
static std::pair<dim3, dim3> calc_launch_params(const int ncols_y, const int nrows_x, const int warp_size, const mmvq_parameter_table_id table_id) {
|
||||
const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_y, table_id) - 1) / calc_rows_per_block(ncols_y, table_id);
|
||||
const dim3 block_nums(nblocks, 1, 1);
|
||||
const dim3 block_dims(warp_size, calc_nwarps(ncols_y, table_id), 1);
|
||||
static std::pair<dim3, dim3> calc_launch_params(
|
||||
const int ncols_dst, const int nrows_x, const int nchannels_y, const int nsamples_y,
|
||||
const int warp_size, const mmvq_parameter_table_id table_id) {
|
||||
const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
|
||||
const dim3 block_nums(nblocks, nchannels_y, nsamples_y);
|
||||
const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
|
||||
return {block_nums, block_dims};
|
||||
}
|
||||
|
||||
template <ggml_type type>
|
||||
static void mul_mat_vec_q_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
static void mul_mat_vec_q_switch_ncols_dst(
|
||||
const void * vx, const void * vy, const int32_t * ids, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_dst,
|
||||
const int stride_row_x, const int stride_col_y, const int stride_col_dst,
|
||||
const int nchannels_x, const int nchannels_y, const int nchannels_dst,
|
||||
const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
|
||||
const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
|
||||
cudaStream_t stream) {
|
||||
|
||||
GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
|
||||
GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
|
||||
GGML_ASSERT(ncols_dst <= MMVQ_MAX_BATCH_SIZE);
|
||||
|
||||
const int channel_ratio = nchannels_dst / nchannels_x;
|
||||
const int sample_ratio = nsamples_dst / nsamples_x;
|
||||
|
||||
const int device = ggml_cuda_get_device();
|
||||
const int warp_size = ggml_cuda_info().devices[device].warp_size;
|
||||
const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
|
||||
|
||||
switch (ncols_y) {
|
||||
GGML_ASSERT(!ids || ncols_dst == 1);
|
||||
switch (ncols_dst) {
|
||||
case 1:
|
||||
{
|
||||
constexpr int c_ncols_y = 1;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
||||
constexpr int c_ncols_dst = 1;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
|
||||
(vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
|
||||
channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
|
||||
break;
|
||||
}
|
||||
case 2:
|
||||
{
|
||||
constexpr int c_ncols_y = 2;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
||||
constexpr int c_ncols_dst = 2;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
|
||||
(vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
|
||||
channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
|
||||
break;
|
||||
}
|
||||
case 3:
|
||||
{
|
||||
constexpr int c_ncols_y = 3;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
||||
constexpr int c_ncols_dst = 3;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
|
||||
(vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
|
||||
channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
|
||||
break;
|
||||
}
|
||||
case 4:
|
||||
{
|
||||
constexpr int c_ncols_y = 4;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
||||
constexpr int c_ncols_dst = 4;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
|
||||
(vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
|
||||
channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
|
||||
break;
|
||||
}
|
||||
case 5:
|
||||
{
|
||||
constexpr int c_ncols_y = 5;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
||||
constexpr int c_ncols_dst = 5;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
|
||||
(vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
|
||||
channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
|
||||
break;
|
||||
}
|
||||
case 6:
|
||||
{
|
||||
constexpr int c_ncols_y = 6;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
||||
constexpr int c_ncols_dst = 6;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
|
||||
(vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
|
||||
channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
|
||||
break;
|
||||
}
|
||||
case 7:
|
||||
{
|
||||
constexpr int c_ncols_y = 7;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
||||
constexpr int c_ncols_dst = 7;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
|
||||
(vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
|
||||
channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
|
||||
break;
|
||||
}
|
||||
case 8:
|
||||
{
|
||||
constexpr int c_ncols_y = 8;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
|
||||
constexpr int c_ncols_dst = 8;
|
||||
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
|
||||
mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
|
||||
(vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
|
||||
channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
|
||||
break;
|
||||
}
|
||||
default:
|
||||
@@ -287,137 +336,213 @@ static void mul_mat_vec_q_cuda(
|
||||
}
|
||||
}
|
||||
|
||||
static void mul_mat_vec_q4_0_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_Q4_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
static void mul_mat_vec_q_switch_type(
|
||||
const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int ncols_dst,
|
||||
const int stride_row_x, const int stride_col_y, const int stride_col_dst,
|
||||
const int nchannels_x, const int nchannels_y, const int nchannels_dst,
|
||||
const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
|
||||
const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
|
||||
cudaStream_t stream) {
|
||||
switch (type_x) {
|
||||
case GGML_TYPE_Q4_0:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_Q4_1:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_Q5_0:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_Q5_1:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_Q8_0:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_Q2_K:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_Q3_K:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_Q4_K:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_Q5_K:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_Q6_K:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ2_XXS:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ2_XS:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ2_S:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ3_XXS:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ1_S:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ1_M:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ4_NL:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ4_XS:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ3_S:
|
||||
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
|
||||
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
|
||||
nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
|
||||
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
|
||||
stream);
|
||||
break;
|
||||
default:
|
||||
GGML_ABORT("fatal error");
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
static void mul_mat_vec_q4_1_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
void ggml_cuda_mul_mat_vec_q(
|
||||
ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
|
||||
GGML_ASSERT( src1->type == GGML_TYPE_F32);
|
||||
GGML_ASSERT( dst->type == GGML_TYPE_F32);
|
||||
GGML_ASSERT(!ids || ids->type == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_Q4_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
GGML_TENSOR_BINARY_OP_LOCALS;
|
||||
|
||||
static void mul_mat_vec_q5_0_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
cudaStream_t stream = ctx.stream();
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_Q5_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
const size_t ts_src0 = ggml_type_size(src0->type);
|
||||
const size_t ts_src1 = ggml_type_size(src1->type);
|
||||
const size_t ts_dst = ggml_type_size(dst->type);
|
||||
|
||||
static void mul_mat_vec_q5_1_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
GGML_ASSERT( nb00 == ts_src0);
|
||||
GGML_ASSERT( nb10 == ts_src1);
|
||||
GGML_ASSERT( nb0 == ts_dst);
|
||||
GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_Q5_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
|
||||
|
||||
static void mul_mat_vec_q8_0_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
const float * src1_d = (const float *) src1->data;
|
||||
const int32_t * ids_d = ids ? (const int32_t *) ids->data : nullptr;
|
||||
float * dst_d = (float *) dst->data;
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_Q8_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
|
||||
ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1);
|
||||
{
|
||||
const int64_t s11 = src1->nb[1] / ts_src1;
|
||||
const int64_t s12 = src1->nb[2] / ts_src1;
|
||||
const int64_t s13 = src1->nb[3] / ts_src1;
|
||||
quantize_row_q8_1_cuda(src1_d, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_q2_K_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
const int64_t s01 = src0->nb[1] / ts_src0;
|
||||
const int64_t s11 = ne10_padded / QK8_1;
|
||||
const int64_t s1 = dst->nb[1] / ts_dst;
|
||||
const int64_t s02 = src0->nb[2] / ts_src0;
|
||||
const int64_t s2 = dst->nb[2] / ts_dst;
|
||||
const int64_t s03 = src0->nb[3] / ts_src0;
|
||||
const int64_t s3 = dst->nb[3] / ts_dst;
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_Q2_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
const int64_t s12 = ne11*s11;
|
||||
const int64_t s13 = ne12*s12;
|
||||
|
||||
static void mul_mat_vec_q3_K_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
// For MUL_MAT_ID the memory layout is different than for MUL_MAT:
|
||||
const int64_t ncols_dst = ids ? ne2 : ne1;
|
||||
const int64_t nchannels_y = ids ? ne11 : ne12;
|
||||
const int64_t nchannels_dst = ids ? ne1 : ne2;
|
||||
const int64_t stride_col_dst = ids ? s2 : s1;
|
||||
const int64_t stride_col_y = ids ? s12 : s11;
|
||||
const int64_t stride_channel_dst = ids ? s1 : s2;
|
||||
const int64_t stride_channel_y = ids ? s11 : s12;
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_Q3_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_q4_K_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_Q4_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_q5_K_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_Q5_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_q6_K_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_Q6_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_iq2_xxs_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_iq2_xs_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_iq2_s_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_IQ2_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_iq3_xxs_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_IQ3_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_iq1_s_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_IQ1_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_iq1_m_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_IQ1_M>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_iq4_nl_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_IQ4_NL>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_iq4_xs_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_IQ4_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
}
|
||||
|
||||
static void mul_mat_vec_iq3_s_q8_1_cuda(
|
||||
const void * vx, const void * vy, float * dst,
|
||||
const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
|
||||
|
||||
mul_mat_vec_q_cuda<GGML_TYPE_IQ3_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
|
||||
mul_mat_vec_q_switch_type(
|
||||
src0->data, src0->type, src1_q8_1.get(), ids_d, dst_d, ne00,
|
||||
ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
|
||||
ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
|
||||
ne03, ne3, s03, s13, s3, stream);
|
||||
}
|
||||
|
||||
void ggml_cuda_op_mul_mat_vec_q(
|
||||
@@ -440,68 +565,12 @@ void ggml_cuda_op_mul_mat_vec_q(
|
||||
// nrows_dst == nrows of the matrix that the kernel writes into
|
||||
const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
|
||||
|
||||
switch (src0->type) {
|
||||
case GGML_TYPE_Q4_0:
|
||||
mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_Q4_1:
|
||||
mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_Q5_0:
|
||||
mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_Q5_1:
|
||||
mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_Q8_0:
|
||||
mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_Q2_K:
|
||||
mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_Q3_K:
|
||||
mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_Q4_K:
|
||||
mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_Q5_K:
|
||||
mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_Q6_K:
|
||||
mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ2_XXS:
|
||||
mul_mat_vec_iq2_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ2_XS:
|
||||
mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ2_S:
|
||||
mul_mat_vec_iq2_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ3_XXS:
|
||||
mul_mat_vec_iq3_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ1_S:
|
||||
mul_mat_vec_iq1_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ1_M:
|
||||
mul_mat_vec_iq1_m_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ4_NL:
|
||||
mul_mat_vec_iq4_nl_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ4_XS:
|
||||
mul_mat_vec_iq4_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
case GGML_TYPE_IQ3_S:
|
||||
mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
|
||||
break;
|
||||
default:
|
||||
GGML_ABORT("fatal error");
|
||||
break;
|
||||
}
|
||||
const int stride_row_x = ne00 / ggml_blck_size(src0->type);
|
||||
const int stride_col_y = src1_padded_row_size / QK8_1;
|
||||
|
||||
mul_mat_vec_q_switch_type(
|
||||
src0_dd_i, src0->type, src1_ddq_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
|
||||
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
|
||||
|
||||
GGML_UNUSED(src1);
|
||||
GGML_UNUSED(dst);
|
||||
|
||||
Reference in New Issue
Block a user