CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-15 11:17:31 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: General GEMV fusion (#16715)

Browse Source
This commit is contained in:
Aman Gupta
2025-10-26 19:28:04 +08:00
committed by GitHub
parent 3cfa9c3f12
commit f77c13b91f
11 changed files with 1096 additions and 166 deletions
Download Patch File Download Diff File Expand all files Collapse all files

314
ggml/src/ggml-cuda/mmvq.cu
View File

@@ -1,5 +1,6 @@
#include "mmvq.cuh"
#include "quantize.cuh"
#include "unary.cuh"
#include "vecdotq.cuh"
#include <cstdint>
@@ -82,7 +83,7 @@ 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_dst, 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_dst) {
case 1:
@@ -136,11 +137,11 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
return 1;
}
template <ggml_type type, int ncols_dst>
// tell the compiler to use as many registers as it wants, see nwarps definition below
template <ggml_type type, int ncols_dst, bool has_fusion>
__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, const int32_t * __restrict__ ids, float * __restrict__ dst,
const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
@@ -169,8 +170,38 @@ static __global__ void mul_mat_vec_q(
const uint32_t sample_x = fastdiv(sample_dst, sample_ratio);
const uint32_t sample_y = sample_dst;
bool use_gate = false;
bool use_bias = false;
bool use_gate_bias = false;
const void * vgate = nullptr;
const float * x_bias = nullptr;
const float * gate_bias = nullptr;
ggml_glu_op active_glu;
if constexpr (has_fusion) {
use_gate = fusion.gate != nullptr;
use_bias = fusion.x_bias != nullptr;
use_gate_bias = fusion.gate_bias != nullptr && use_gate;
vgate = fusion.gate;
x_bias = (const float *) fusion.x_bias;
gate_bias = (const float *) fusion.gate_bias;
active_glu = fusion.glu_op;
}
const uint32_t channel_bias = ids ? channel_x : channel_dst;
if constexpr (has_fusion) {
if (use_bias) {
x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
}
if (use_gate_bias) {
gate_bias = gate_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
}
}
// partial sum for each thread
float tmp[ncols_dst][rows_per_cuda_block] = {{0.0f}};
float tmp_gate[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;
@@ -187,17 +218,35 @@ static __global__ void mul_mat_vec_q(
for (int i = 0; i < rows_per_cuda_block; ++i) {
tmp[j][i] += vec_dot_q_cuda(
vx, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
if constexpr (has_fusion) {
if (use_gate) {
tmp_gate[j][i] += vec_dot_q_cuda(
vgate, &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_dst][rows_per_cuda_block][warp_size];
__shared__ float tmp_shared_gate[(has_fusion && (nwarps-1 > 0)) ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
if constexpr (!has_fusion) {
(void) tmp_shared_gate;
} else if (!use_gate) {
(void) tmp_shared_gate;
}
if (threadIdx.y > 0) {
#pragma unroll
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];
if constexpr (has_fusion) {
if (use_gate) {
tmp_shared_gate[threadIdx.y-1][j][i][threadIdx.x] = tmp_gate[j][i];
}
}
}
}
}
@@ -216,12 +265,49 @@ static __global__ void mul_mat_vec_q(
#pragma unroll
for (int l = 0; l < nwarps-1; ++l) {
tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
if constexpr (has_fusion) {
if (use_gate) {
tmp_gate[j][i] += tmp_shared_gate[l][j][i][threadIdx.x];
}
}
}
tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
if constexpr (has_fusion) {
if (use_gate) {
tmp_gate[j][i] = warp_reduce_sum<warp_size>(tmp_gate[j][i]);
}
}
}
if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
dst[j*stride_col_dst + threadIdx.x] = tmp[j][threadIdx.x];
float result = tmp[j][threadIdx.x];
if constexpr (has_fusion) {
if (use_bias) {
result += x_bias[j*stride_col_dst + threadIdx.x];
}
if (use_gate) {
float gate_value = tmp_gate[j][threadIdx.x];
if (use_gate_bias) {
gate_value += gate_bias[j*stride_col_dst + threadIdx.x];
}
switch (active_glu) {
case GGML_GLU_OP_SWIGLU:
result *= ggml_cuda_op_silu_single(gate_value);
break;
case GGML_GLU_OP_GEGLU:
result *= ggml_cuda_op_gelu_single(gate_value);
break;
case GGML_GLU_OP_SWIGLU_OAI: {
result = ggml_cuda_op_swiglu_oai_single(gate_value, result);
break;
}
default:
result = result * gate_value;
break;
}
}
}
dst[j*stride_col_dst + threadIdx.x] = result;
}
}
}
@@ -235,9 +321,37 @@ static std::pair<dim3, dim3> calc_launch_params(
return {block_nums, block_dims};
}
template<ggml_type type, int c_ncols_dst>
static void mul_mat_vec_q_switch_fusion(
const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared, cudaStream_t stream) {
const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
if constexpr (c_ncols_dst == 1) {
if (has_fusion) {
mul_mat_vec_q<type, c_ncols_dst, true><<<block_nums, block_dims, nbytes_shared, stream>>>
(vx, vy, ids, fusion, 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);
return;
}
}
GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
mul_mat_vec_q<type, c_ncols_dst, false><<<block_nums, block_dims, nbytes_shared, stream>>>
(vx, vy, ids, fusion, 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);
}
template <ggml_type type>
static void mul_mat_vec_q_switch_ncols_dst(
const void * vx, const void * vy, const int32_t * ids, float * dst,
const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, 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,
@@ -256,80 +370,83 @@ static void mul_mat_vec_q_switch_ncols_dst(
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);
const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
GGML_ASSERT(!ids || ncols_dst == 1);
switch (ncols_dst) {
case 1: {
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_fd, stride_row_x, stride_col_y, stride_col_dst,
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
dims.first, dims.second, 0, stream);
} break;
case 2: {
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_fd, stride_row_x, stride_col_y, stride_col_dst,
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
dims.first, dims.second, 0, stream);
} break;
case 3: {
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_fd, stride_row_x, stride_col_y, stride_col_dst,
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
dims.first, dims.second, 0, stream);
} break;
case 4: {
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_fd, stride_row_x, stride_col_y, stride_col_dst,
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
dims.first, dims.second, 0, stream);
} break;
case 5: {
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_fd, stride_row_x, stride_col_y, stride_col_dst,
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
dims.first, dims.second, 0, stream);
} break;
case 6: {
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_fd, stride_row_x, stride_col_y, stride_col_dst,
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
dims.first, dims.second, 0, stream);
} break;
case 7: {
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_fd, stride_row_x, stride_col_y, stride_col_dst,
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
dims.first, dims.second, 0, stream);
} break;
case 8: {
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_fd, stride_row_x, stride_col_y, stride_col_dst,
mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst);
sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
dims.first, dims.second, 0, stream);
} break;
default:
GGML_ABORT("fatal error");
break;
}
}
GGML_UNUSED(has_fusion);
}
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 void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, 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,
@@ -339,143 +456,123 @@ static void mul_mat_vec_q_switch_type(
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
break;
case GGML_TYPE_MXFP4:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_MXFP4>
(vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
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,
(vx, vy, ids, fusion, 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);
nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
break;
default:
GGML_ABORT("fatal error");
@@ -484,7 +581,8 @@ static void mul_mat_vec_q_switch_type(
}
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_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
const ggml_cuda_mm_fusion_args_host * fusion) {
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.
@@ -508,6 +606,31 @@ void ggml_cuda_mul_mat_vec_q(
const int32_t * ids_d = ids ? (const int32_t *) ids->data : nullptr;
float * dst_d = (float *) dst->data;
ggml_cuda_mm_fusion_args_device fusion_local{};
if (fusion) {
GGML_ASSERT( !ids || dst->ne[2] == 1);
GGML_ASSERT( ids || dst->ne[1] == 1);
if (fusion->x_bias) {
GGML_ASSERT(fusion->x_bias->type == GGML_TYPE_F32);
GGML_ASSERT(fusion->x_bias->ne[0] == dst->ne[0]);
GGML_ASSERT(!ids || fusion->x_bias->ne[1] == src0->ne[2]);
fusion_local.x_bias = fusion->x_bias->data;
}
if (fusion->gate) {
GGML_ASSERT(fusion->gate->type == src0->type && ggml_are_same_stride(fusion->gate, src0));
fusion_local.gate = fusion->gate->data;
}
if (fusion->gate_bias) {
GGML_ASSERT(fusion->gate_bias->type == GGML_TYPE_F32);
GGML_ASSERT(fusion->gate_bias->ne[0] == dst->ne[0]);
GGML_ASSERT(!ids || fusion->gate_bias->ne[1] == src0->ne[2]);
fusion_local.gate_bias = fusion->gate_bias->data;
}
fusion_local.glu_op = fusion->glu_op;
}
// If src0 is a temporary compute buffer, clear any potential padding.
if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
const size_t size_data = ggml_nbytes(src0);
@@ -549,10 +672,10 @@ void ggml_cuda_mul_mat_vec_q(
const int64_t stride_channel_y = ids ? s11 : s12;
mul_mat_vec_q_switch_type(
src0->data, src0->type, src1_q8_1.get(), ids_d, dst_d, ne00,
src0->data, src0->type, src1_q8_1.get(), ids_d, fusion_local, 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);
ne03, ne3, s03, s13, s3, stream);
}
void ggml_cuda_op_mul_mat_vec_q(
@@ -578,8 +701,9 @@ void ggml_cuda_op_mul_mat_vec_q(
const int stride_row_x = ne00 / ggml_blck_size(src0->type);
const int stride_col_y = src1_padded_row_size / QK8_1;
ggml_cuda_mm_fusion_args_device fusion_local{};
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,
src0_dd_i, src0->type, src1_ddq_i, nullptr, fusion_local, 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_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);