happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

CUDA: use mmvq for mul-mat-id for small batch sizes

This commit is contained in:
Aman Gupta 2026-01-20 13:03:43 +01:00
parent 2634ed207a
commit 1c77ccf01d
2 changed files with 63 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -2279,7 +2279,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
if (ne2 == 1) {
if (ne2 <= MMVQ_MAX_BATCH_SIZE) {
if (ggml_is_quantized(src0->type)) {
ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
} else {

108
ggml/src/ggml-cuda/mmvq.cu
View File

@ -145,7 +145,8 @@ static __global__ void mul_mat_vec_q(
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 uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
const uint32_t ids_stride) {
constexpr int qk = ggml_cuda_type_traits<type>::qk;
constexpr int qi = ggml_cuda_type_traits<type>::qi;
@ -162,11 +163,12 @@ static __global__ void mul_mat_vec_q(
const int blocks_per_row_x = ncols_x / qk;
constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
// The MUL_MAT_ID code path with ids != nullptr is only implemented for ncols_dst == 1.
// for MUL_MAT_ID - blockIdx.y = n_expert_used, blockIdx.z = ncols_dst (tokens)
const uint32_t channel_dst = blockIdx.y;
const uint32_t channel_x = ncols_dst == 1 && ids ? ids[channel_dst] : fastdiv(channel_dst, channel_ratio);
const uint32_t channel_y = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
const uint32_t sample_dst = blockIdx.z;
const uint32_t token_idx = ids ? blockIdx.z : 0;
const uint32_t channel_x = ids ? ids[blockIdx.y + token_idx * ids_stride] : fastdiv(channel_dst, channel_ratio);
const uint32_t channel_y = ids ? fastmodulo(blockIdx.y, nchannels_y) : channel_dst;
const uint32_t sample_dst = ids ? 0 : blockIdx.z;
const uint32_t sample_x = fastdiv(sample_dst, sample_ratio);
const uint32_t sample_y = sample_dst;
@ -221,7 +223,7 @@ static __global__ void mul_mat_vec_q(
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 block_q8_1 * y = ((const block_q8_1 *) vy) + token_idx*stride_col_y + 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) {
@ -273,7 +275,7 @@ static __global__ void mul_mat_vec_q(
return;
}
dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;
dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + token_idx*stride_col_dst + row0;
// sum up partial sums and write back result
#pragma unroll
@ -335,10 +337,10 @@ static __global__ void mul_mat_vec_q(
}
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 ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
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_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
return {block_nums, block_dims};
}
@ -350,7 +352,8 @@ static void mul_mat_vec_q_switch_fusion(
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 dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared,
const uint32_t ids_stride, cudaStream_t stream) {
const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
if constexpr (c_ncols_dst == 1) {
@ -358,7 +361,7 @@ static void mul_mat_vec_q_switch_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);
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
return;
}
}
@ -368,7 +371,7 @@ static void mul_mat_vec_q_switch_fusion(
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);
sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
}
template <ggml_type type>
@ -379,7 +382,7 @@ static void mul_mat_vec_q_switch_ncols_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) {
const int ids_stride, cudaStream_t stream) {
GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
GGML_ASSERT(ncols_dst <= MMVQ_MAX_BATCH_SIZE);
@ -393,8 +396,19 @@ static void mul_mat_vec_q_switch_ncols_dst(
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;
const bool has_ids = ids != nullptr;
if (has_ids) {
// note: batching ncols_dst is not possible because token use different experts, so we use ncols_dst = 1 and iterate via blockIdx.z
constexpr int c_ncols_dst = 1;
std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
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,
dims.first, dims.second, 0, ids_stride, stream);
return;
}
GGML_ASSERT(!ids || ncols_dst == 1);
switch (ncols_dst) {
case 1: {
constexpr int c_ncols_dst = 1;
@ -402,7 +416,7 @@ static void mul_mat_vec_q_switch_ncols_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,
dims.first, dims.second, 0, stream);
dims.first, dims.second, 0, ids_stride, stream);
} break;
case 2: {
constexpr int c_ncols_dst = 2;
@ -410,7 +424,7 @@ static void mul_mat_vec_q_switch_ncols_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,
dims.first, dims.second, 0, stream);
dims.first, dims.second, 0, ids_stride, stream);
} break;
case 3: {
constexpr int c_ncols_dst = 3;
@ -418,7 +432,7 @@ static void mul_mat_vec_q_switch_ncols_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,
dims.first, dims.second, 0, stream);
dims.first, dims.second, 0, ids_stride, stream);
} break;
case 4: {
constexpr int c_ncols_dst = 4;
@ -426,7 +440,7 @@ static void mul_mat_vec_q_switch_ncols_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,
dims.first, dims.second, 0, stream);
dims.first, dims.second, 0, ids_stride, stream);
} break;
case 5: {
constexpr int c_ncols_dst = 5;
@ -434,7 +448,7 @@ static void mul_mat_vec_q_switch_ncols_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,
dims.first, dims.second, 0, stream);
dims.first, dims.second, 0, ids_stride, stream);
} break;
case 6: {
constexpr int c_ncols_dst = 6;
@ -442,7 +456,7 @@ static void mul_mat_vec_q_switch_ncols_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,
dims.first, dims.second, 0, stream);
dims.first, dims.second, 0, ids_stride, stream);
} break;
case 7: {
constexpr int c_ncols_dst = 7;
@ -450,7 +464,7 @@ static void mul_mat_vec_q_switch_ncols_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,
dims.first, dims.second, 0, stream);
dims.first, dims.second, 0, ids_stride, stream);
} break;
case 8: {
constexpr int c_ncols_dst = 8;
@ -458,7 +472,7 @@ static void mul_mat_vec_q_switch_ncols_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,
dims.first, dims.second, 0, stream);
dims.first, dims.second, 0, ids_stride, stream);
} break;
default:
GGML_ABORT("fatal error");
@ -474,127 +488,127 @@ static void mul_mat_vec_q_switch_type(
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) {
const int ids_stride, 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, 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, ids_stride, stream);
break;
case GGML_TYPE_Q4_1:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
(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, ids_stride, stream);
break;
case GGML_TYPE_Q5_0:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
(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, ids_stride, stream);
break;
case GGML_TYPE_Q5_1:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
(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, ids_stride, stream);
break;
case GGML_TYPE_Q8_0:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
(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, ids_stride, stream);
break;
case GGML_TYPE_MXFP4:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_MXFP4>
(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, ids_stride, stream);
break;
case GGML_TYPE_Q2_K:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
(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, ids_stride, stream);
break;
case GGML_TYPE_Q3_K:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
(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, ids_stride, stream);
break;
case GGML_TYPE_Q4_K:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
(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, ids_stride, stream);
break;
case GGML_TYPE_Q5_K:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
(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, ids_stride, stream);
break;
case GGML_TYPE_Q6_K:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
(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, ids_stride, stream);
break;
case GGML_TYPE_IQ2_XXS:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
(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, ids_stride, stream);
break;
case GGML_TYPE_IQ2_XS:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
(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, ids_stride, stream);
break;
case GGML_TYPE_IQ2_S:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
(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, ids_stride, stream);
break;
case GGML_TYPE_IQ3_XXS:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
(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, ids_stride, stream);
break;
case GGML_TYPE_IQ1_S:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
(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, ids_stride, stream);
break;
case GGML_TYPE_IQ1_M:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
(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, ids_stride, stream);
break;
case GGML_TYPE_IQ4_NL:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
(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, ids_stride, stream);
break;
case GGML_TYPE_IQ4_XS:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
(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, ids_stride, stream);
break;
case GGML_TYPE_IQ3_S:
mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
(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, ids_stride, stream);
break;
default:
GGML_ABORT("fatal error");
@ -622,7 +636,7 @@ void ggml_cuda_mul_mat_vec_q(
GGML_ASSERT( nb0 == ts_dst);
GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
GGML_ASSERT(!ids || ne12 <= MMVQ_MAX_BATCH_SIZE);
const float * src1_d = (const float *) src1->data;
const int32_t * ids_d = ids ? (const int32_t *) ids->data : nullptr;
@ -693,11 +707,13 @@ void ggml_cuda_mul_mat_vec_q(
const int64_t stride_channel_dst = ids ? s1 : s2;
const int64_t stride_channel_y = ids ? s11 : s12;
const int64_t ids_stride = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
mul_mat_vec_q_switch_type(
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, ids_stride, stream);
}
void ggml_cuda_op_mul_mat_vec_q(
@ -726,7 +742,7 @@ void ggml_cuda_op_mul_mat_vec_q(
ggml_cuda_mm_fusion_args_device fusion_local{};
mul_mat_vec_q_switch_type(
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);
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, stream);
GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);
}