ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-22 23:49:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

MMQ for iq2_ks

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-05-14 18:41:42 +03:00
parent 1aceccae35
commit f069a57817
3 changed files with 99 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-cuda/mmq.cu
View File

@@ -94,6 +94,9 @@ void ggml_cuda_op_mul_mat_q(
case GGML_TYPE_IQ4_KS:
mul_mat_q_case<GGML_TYPE_IQ4_KS>(ctx, args, stream);
break;
case GGML_TYPE_IQ2_KS:
mul_mat_q_case<GGML_TYPE_IQ2_KS>(ctx, args, stream);
break;
case GGML_TYPE_IQ2_K:
mul_mat_q_case<GGML_TYPE_IQ2_K>(ctx, args, stream);
break;
@@ -147,6 +150,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_KS:
case GGML_TYPE_IQ2_KS:
case GGML_TYPE_IQ2_K:
case GGML_TYPE_IQ3_K:
case GGML_TYPE_IQ4_K:

90
ggml/src/ggml-cuda/mmq.cuh
View File

@@ -82,6 +82,7 @@ static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
return MMQ_Q8_1_DS_LAYOUT_DS4;
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ2_KS:
case GGML_TYPE_IQ2_K:
case GGML_TYPE_IQ3_K:
case GGML_TYPE_IQ4_KS:
@@ -186,6 +187,7 @@ static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml
case GGML_TYPE_IQ4_XS : return MMQ_DP4A_TXS_Q8_0;
case GGML_TYPE_IQ4_NL : return MMQ_DP4A_TXS_Q8_0;
case GGML_TYPE_IQ4_KS : return MMQ_DP4A_TXS_Q8_0;
case GGML_TYPE_IQ2_KS : return MMQ_DP4A_TXS_Q8_0;
case GGML_TYPE_IQ2_K : return MMQ_DP4A_TXS_Q8_0_16;
case GGML_TYPE_IQ3_K : return MMQ_DP4A_TXS_Q8_0_16;
case GGML_TYPE_IQ4_K : return MMQ_DP4A_TXS_Q8_0_16;
@@ -229,6 +231,7 @@ static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
case GGML_TYPE_IQ4_XS : return MMQ_MMA_TILE_X_K_Q8_0;
case GGML_TYPE_IQ4_NL : return MMQ_MMA_TILE_X_K_Q8_0;
case GGML_TYPE_IQ4_KS : return MMQ_MMA_TILE_X_K_Q8_0;
case GGML_TYPE_IQ2_KS : return MMQ_MMA_TILE_X_K_Q8_0;
case GGML_TYPE_IQ2_K : return MMQ_MMA_TILE_X_K_Q3_K;
case GGML_TYPE_IQ3_K : return MMQ_MMA_TILE_X_K_Q3_K;
case GGML_TYPE_IQ4_K : return MMQ_MMA_TILE_X_K_Q3_K;
@@ -2373,6 +2376,84 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
}
}
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_ks(
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
#ifdef INT8_MMA_AVAILABLE
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + WARP_SIZE*2);
#else
constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, mmq_y);
int * x_qs = (int *) x_tile;
float * x_df = (float *) (x_qs + txs.qs);
#endif // INT8_MMA_AVAILABLE
const int kqsx = threadIdx.x%16;
auto values = iq2nl_values;
uint32_t aux32[4];
const uint8_t * aux8 = (const uint8_t *)aux32;
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += 2*nwarps) {
int i = i0 + 2*threadIdx.y + threadIdx.x/16;
if (need_check) {
i = min(i, i_max);
}
const block_iq2_ks * bxi = (const block_iq2_ks *)(x + i*stride + sizeof(half)) + kbx0;
uint16_t extra = bxi->extra >> 8*(kqsx/8);
int q2 = get_int_b2(bxi->qs, kqsx);
aux32[0] = ((q2 >> 0) & 0x03030303) | (((extra << 2) & 4) * 0x01010101);
aux32[1] = ((q2 >> 2) & 0x03030303) | (((extra << 1) & 4) * 0x01010101);
aux32[2] = ((q2 >> 4) & 0x03030303) | (((extra >> 0) & 4) * 0x01010101);
aux32[3] = ((q2 >> 6) & 0x03030303) | (((extra >> 1) & 4) * 0x01010101);
const char4 val0 = make_char4(values[aux8[ 0]], values[aux8[ 1]], values[aux8[ 2]], values[aux8[ 3]]);
const char4 val1 = make_char4(values[aux8[ 4]], values[aux8[ 5]], values[aux8[ 6]], values[aux8[ 7]]);
const char4 val2 = make_char4(values[aux8[ 8]], values[aux8[ 9]], values[aux8[10]], values[aux8[11]]);
const char4 val3 = make_char4(values[aux8[12]], values[aux8[13]], values[aux8[14]], values[aux8[15]]);
#ifdef INT8_MMA_AVAILABLE
x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx%8 + 32*(kqsx/8) + 0] = *(const int *)&val0;
x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx%8 + 32*(kqsx/8) + 8] = *(const int *)&val1;
x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx%8 + 32*(kqsx/8) + 16] = *(const int *)&val2;
x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx%8 + 32*(kqsx/8) + 24] = *(const int *)&val3;
#else
x_qs[i*(2*WARP_SIZE + 1) + kqsx%8 + 32*(kqsx/8) + 0] = *(const int *)&val0;
x_qs[i*(2*WARP_SIZE + 1) + kqsx%8 + 32*(kqsx/8) + 8] = *(const int *)&val1;
x_qs[i*(2*WARP_SIZE + 1) + kqsx%8 + 32*(kqsx/8) + 16] = *(const int *)&val2;
x_qs[i*(2*WARP_SIZE + 1) + kqsx%8 + 32*(kqsx/8) + 24] = *(const int *)&val3;
#endif // INT8_MMA_AVAILABLE
}
#pragma unroll
for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
int i = i0 + threadIdx.y * 8 + threadIdx.x / 4;
if (need_check) {
i = min(i, i_max);
}
const half * dptr = (const half *)(x + i*stride);
const float d = dptr[0];
const block_iq2_ks * bxi = (const block_iq2_ks *)(dptr + 1) + kbx0;
const int ls1 = ((bxi->scales[threadIdx.x % 4] >> 0) & 0xf) | ((bxi->extra >> (4 + 2*(threadIdx.x % 4))) & 0x10);
const int ls2 = ((bxi->scales[threadIdx.x % 4] >> 4) & 0xf) | ((bxi->extra >> (5 + 2*(threadIdx.x % 4))) & 0x10);
#ifdef INT8_MMA_AVAILABLE
x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + 2*(threadIdx.x % 4) + 0] = d * (ls1 - 16);
x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + 2*(threadIdx.x % 4) + 1] = d * (ls2 - 16);
#else
x_df[i*(WARP_SIZE/4) + i/4 + 2*(threadIdx.x % 4) + 0] = d * (ls1 - 16);
x_df[i*(WARP_SIZE/4) + i/4 + 2*(threadIdx.x % 4) + 1] = d * (ls2 - 16);
#endif // INT8_MMA_AVAILABLE
}
}
template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_k(
const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
@@ -3042,6 +3123,14 @@ struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ6_K> {
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ2_KS> {
static constexpr int vdr = VDR_IQ4_XS_Q8_1_MMQ;
static constexpr load_tiles_mmq_t load_tiles = load_tiles_iq2_ks<mmq_y, nwarps, need_check>;
static constexpr vec_dot_mmq_t vec_dot_mma = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
static constexpr vec_dot_mmq_t vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
};
template <int mmq_x, int mmq_y, int nwarps, bool need_check>
struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_KS> {
static constexpr int vdr = VDR_IQ4_XS_Q8_1_MMQ;
@@ -3487,6 +3576,7 @@ extern DECL_MMQ_CASE(GGML_TYPE_IQ1_S);
extern DECL_MMQ_CASE(GGML_TYPE_IQ4_NL);
extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS);
extern DECL_MMQ_CASE(GGML_TYPE_IQ4_KS);
extern DECL_MMQ_CASE(GGML_TYPE_IQ2_KS);
extern DECL_MMQ_CASE(GGML_TYPE_IQ2_K);
extern DECL_MMQ_CASE(GGML_TYPE_IQ3_K);
extern DECL_MMQ_CASE(GGML_TYPE_IQ4_K);

5
ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_ks.cu Normal file
View File

@@ -0,0 +1,5 @@
// This file has been autogenerated by generate_cu_files.py, do not edit manually.
#include "../mmq.cuh"
DECL_MMQ_CASE(GGML_TYPE_IQ2_KS);