ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-24 07:04:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

New iq4_kt trellis

Browse Source 
The new trellis generates int8_t values via
sum_as_uint8_t[(ka * idx + kb) & 0x3f33f3f3f] - 126.
CUDA dequantize works.
AVX2 case Ny > 32 works, and we get 273 t/s for L3-8B.
PPL is on par or even slightly lower than original QTIP trellis.
This commit is contained in:
Iwan Kawrakow
2025-06-07 12:30:37 +03:00
parent c410cc72bb
commit e558992f0c
5 changed files with 181 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-cuda/convert.cu
View File

@@ -343,13 +343,9 @@ inline __device__ int nearest_int(float fval) {
float __device__ __forceinline__ trellis_next(uint32_t& val) {
constexpr uint32_t ka = 89226354;
constexpr uint32_t kb = 64248484;
constexpr uint32_t kmask = 0x8fff8fff;
constexpr uint32_t km32 = 0x3b603b60;
uint32_t s;
const half * h = (const half *)&s;
val = ka*val + kb;
s = (val & kmask) ^ km32;
return (float)(h[0]+h[1]);
//return ggml_cuda_dp4a(val & 0x3f3f3f3f, 0x01010101, 0x82828282);
return ggml_cuda_dp4a(val & 0x3f3f3f3f, 0x01010101, -126);
}
template<typename dst_t>
@@ -367,7 +363,7 @@ static __global__ void dequantize_block_iq2_kt(const void * __restrict__ vx, dst
dst_t * y = yy + ii*QK_K + 8*ib;
const uint16_t * ql = (const uint16_t *)x[i].ql;
uint32_t idx = ql[ib] + 4096;
const float dl = scale * iq4k_values[((x[i].scales[(ib/4)%4] >> 4*(ib/16)) & 0xf)] * 31.75f * 1.05f;
const float dl = scale * iq4k_values[((x[i].scales[(ib/4)%4] >> 4*(ib/16)) & 0xf)] * 1.05f;
for (int j = 0; j < 8; ++j) {
y[j] = dl * trellis_next(idx);
}
@@ -401,7 +397,7 @@ static __global__ void dequantize_block_iq4_kt(const void * __restrict__ vx, dst
int64_t ii = blockIdx.x;
int64_t row = (QK_K * ii) / n_per_row;
const float * dptr = (const float *)((const char *)vx + row * row_size);
float scale = dptr[0] * 31.75f * 1.01f;
float scale = dptr[0] * 1.00f;
float row_av = dptr[1];
const block_iq4_kt * x = (const block_iq4_kt *)(dptr + 2);
const int64_t i = ii - (row*n_per_row)/QK_K;

11
ggml/src/ggml.c
View File

@@ -1630,11 +1630,16 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.from_float = quantize_row_iq4_kt,
.from_float_ref = (ggml_from_float_t)quantize_row_iq4_kt_ref,
.vec_dot = vec_dot_iq4_kt_q8_k,
#ifdef __ARM_NEON
.vec_dot_type = GGML_TYPE_F16,
#if defined __AVX2__
.vec_dot_type = GGML_TYPE_Q8_2_X4,
#else
.vec_dot_type = GGML_TYPE_F32,
.vec_dot_type = GGML_TYPE_Q8_0_X4,
#endif
//#ifdef __ARM_NEON
// .vec_dot_type = GGML_TYPE_F16,
//#else
// .vec_dot_type = GGML_TYPE_F32,
//#endif
.nrows = 1,
.row_meta_size = 8,
},

157
ggml/src/iqk/iqk_gemm_ktquants.cpp
View File

@@ -97,6 +97,43 @@ struct Trellis2 {
}
};
struct Trellis3 {
constexpr static uint32_t ka = 89226354;
constexpr static uint32_t kb = 64248484;
constexpr static uint32_t ka1 = ka*ka;
constexpr static uint32_t kb1 = kb*ka+kb;
constexpr static uint32_t ka2 = ka1*ka;
constexpr static uint32_t kb2 = kb1*ka+kb;
constexpr static uint32_t ka3 = ka2*ka;
constexpr static uint32_t kb3 = kb2*ka+kb;
const __m256i mka = _mm256_setr_epi32(ka, ka1, ka2, ka3, ka, ka1, ka2, ka3);
const __m256i mkb = _mm256_setr_epi32(kb, kb1, kb2, kb3, kb, kb1, kb2, kb3);
const __m256i shuffle = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
inline __m256i next8(uint32_t val1, uint32_t val2) const {
__m256i mval = MM256_SET_M128I(_mm_set1_epi32(val2), _mm_set1_epi32(val1));
return _mm256_add_epi32(_mm256_mullo_epi32(mval, mka), mkb);
}
inline __m256 gen8(uint32_t val1, uint32_t val2) const {
auto v8 = _mm256_and_si256(next8(val1, val2), _mm256_set1_epi32(0x3f3f3f3f));
auto i8 = _mm256_dpbusd_epi32(_mm256_set1_epi32(-126), _mm256_set1_epi32(0x01010101), v8);
return _mm256_cvtepi32_ps(i8);
}
inline __m256i next32(const uint32_t * val) const {
__m256i aux[4];
for (int i = 0; i < 4; ++i) {
auto i8 = _mm256_and_si256(next8(val[2*i+0], val[2*i+1]), _mm256_set1_epi32(0x3f3f3f3f));
aux[i] = _mm256_dpbusd_epi32(_mm256_set1_epi32(-126), _mm256_set1_epi32(0x01010101), i8);
}
aux[0] = _mm256_packs_epi32(aux[0], aux[1]); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15
aux[2] = _mm256_packs_epi32(aux[2], aux[3]); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31
aux[0] = _mm256_packs_epi16(aux[0], aux[2]); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27
// 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
return _mm256_permutevar8x32_epi32(aux[0], shuffle);
}
};
void iqk_dequantize_iq2_kt(int n, const void * vx, size_t bx, float * y, size_t stride_y, int nrc_x) {
GGML_ASSERT(n%QK_K == 0);
const int nb = n/QK_K;
@@ -315,19 +352,121 @@ void mul_mat_iq3_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& inf
}
}
// Q8_0 repacking:
// for (int ib = 0; ib < nblock; ++ib) {
// for (int k = 0; k < 8; ++k) y[ib].d[k] = x8[k][ib].d;
// for (int l = 0; l < 4; ++l) {
// for (int k = 0; k < 8; ++k) for (int i = 0; i < 4; ++i) {
// y[ib].qs[32*l+4*k+i+ 0] = x8[k][ib].qs[i+4*l+ 0];
// y[ib].qs[32*l+4*k+i+128] = x8[k][ib].qs[i+4*l+16];
// as uint32_t
// y[ib].qs[8*l+k+ 0] = x8[k][ib].qs[l+ 0];
// y[ib].qs[8*l+k+32] = x8[k][ib].qs[l+16];
// }
// }
// }
void iqk_dequantize_iq4_kt_q80_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
GGML_ASSERT(n%QK_K == 0);
GGML_ASSERT(nrc_x%8 == 0);
const int nb = n/QK_K;
constexpr int kNumGroups = 64;
Trellis3 trellis;
block_q8_0_r8 * y = (block_q8_0_r8 *)vy;
const block_iq4_kt * x8[8];
float dkt[8];
int32_t ls[8];
uint32_t idx0[8], idx[16];
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) {
const float * dptr = (const float *)((const char*)vx + (ix+k)*bx);
dkt[k] = dptr[0];
x8[k] = (const block_iq4_kt *)(dptr + 2);
}
auto vd = _mm256_loadu_ps(dkt);
for (int i = 0; i < nb; ++i) {
for (int ib = 0; ib < QK_K/32; ++ib) {
for (int k = 0; k < 8; ++k) {
ls[k] = ((x8[k][i].qs[ib] & 0xff) >> 1) - 64;
idx0[k] = ((x8[k][i].qs[ib] & 1) << 15) + 4096;
}
auto scales = _mm256_mul_ps(vd, _mm256_cvtepi32_ps(_mm256_loadu_si256((const __m256i *)ls)));
_mm_storeu_si128((__m128i *)y[ib].d, _mm256_cvtps_ph(scales, _MM_FROUND_TO_NEAREST_INT));
//for (int k = 0; k < 8; ++k) {
// auto shb = x8[k][i].qs;
// const uint8_t * ql = (const uint8_t *)(shb + 8);
// const uint8_t * qh = ql + kNumGroups;
// for (int ib = 0; ib < 4; ++ib) {
// uint32_t offset1 = ((shb[ib+0] & 1) << 15) + 4096;
// uint32_t offset2 = ((shb[ib+4] & 1) << 15) + 4096;
// for (int j = 0; j < 4; ++j) {
// const uint32_t sh1 = shb[ib+0] >> (8 + 6*j);
// const uint32_t sh2 = shb[ib+4] >> (8 + 6*j);
// idx[64*ib + 16*j + k ] = ql[8*ib+2*j+ 0] + ((qh[8*ib+2*j+0] << 8) & 0xf00) + ((sh1 & 7) << 12) + offset1;
// idx[64*ib + 16*j + k + 8] = ql[8*ib+2*j+ 1] + ((qh[8*ib+2*j+1] << 8) & 0xf00) + ((sh1 & 56) << 9) + offset1;
// idx[64*ib + 16*j + k + 256] = ql[8*ib+2*j+32] + ((qh[8*ib+2*j+0] << 4) & 0xf00) + ((sh2 & 7) << 12) + offset2;
// idx[64*ib + 16*j + k + 264] = ql[8*ib+2*j+33] + ((qh[8*ib+2*j+1] << 4) & 0xf00) + ((sh2 & 56) << 9) + offset2;
// //uint32_t val1 = ql[8*ib+2*j+ 0] + ((qh[8*ib+2*j+0] << 8) & 0xf00) + ((sh1 & 7) << 12) + offset1;
// //uint32_t val2 = ql[8*ib+2*j+32] + ((qh[8*ib+2*j+0] << 4) & 0xf00) + ((sh2 & 7) << 12) + offset2;
// //uint32_t val3 = ql[8*ib+2*j+ 1] + ((qh[8*ib+2*j+1] << 8) & 0xf00) + ((sh1 & 56) << 9) + offset1;
// //uint32_t val4 = ql[8*ib+2*j+33] + ((qh[8*ib+2*j+1] << 4) & 0xf00) + ((sh2 & 56) << 9) + offset2;
// //auto x_val1 = _mm256_fmadd_ps(scale1, trellis.gen8(val1, val3), dav);
// //auto x_val2 = _mm256_fmadd_ps(scale2, trellis.gen8(val2, val4), dav);
// //_mm256_storeu_ps(y + i*QK_K + 32*ib + 8*j, x_val1);
// //_mm256_storeu_ps(y + i*QK_K + 32*ib + 8*j + QK_K/2, x_val2);
// }
// }
//}
//for (int j = 0; j < 64; ++j) {
// _mm256_storeu_si256((__m256i *)y[j/8].qs+(j%8), trellis.next32(idx+8*j));
//}
//int shift1 = 8 - 4*(ib/4);
//for (int j = 0; j < 4; ++j) {
// for (int k = 0; k < 8; ++k) {
// const uint8_t * ql = (const uint8_t *)(x8[k][i].qs + 8);
// const uint8_t * qh = ql + kNumGroups;
// const uint32_t sh = x8[k][i].qs[ib] >> (8 + 6*j);
// idx[k+0] = ql[8*ib+2*j+0] + ((qh[8*(ib%4)+2*j+0] << shift1) & 0xf00) + ((sh & 7) << 12) + idx0[k];
// idx[k+8] = ql[8*ib+2*j+1] + ((qh[8*(ib%4)+2*j+1] << shift1) & 0xf00) + ((sh & 56) << 9) + idx0[k];
// }
// _mm256_storeu_si256((__m256i *)y[ib].qs+2*j+0, trellis.next32(idx+0));
// _mm256_storeu_si256((__m256i *)y[ib].qs+2*j+1, trellis.next32(idx+8));
//}
int shift1 = 8 - 4*(ib/4);
for (int j = 0; j < 8; ++j) {
for (int k = 0; k < 8; ++k) {
const uint8_t * ql = (const uint8_t *)(x8[k][i].qs + 8);
const uint8_t * qh = ql + kNumGroups;
const uint32_t sh = x8[k][i].qs[ib] >> (8 + 3*j);
idx[k+0] = ql[8*ib+j] + ((qh[8*(ib%4)+j] << shift1) & 0xf00) + ((sh & 7) << 12) + idx0[k];
}
_mm256_storeu_si256((__m256i *)y[ib].qs+j, trellis.next32(idx));
}
}
y += 8; // = QK_K/32;
}
}
}
void iqk_dequantize_iq4_kt(int n, const void * vx, size_t bx, float * y, size_t stride_y, int nrc_x) {
GGML_ASSERT(n%QK_K == 0);
const int nb = n/QK_K;
constexpr int kNumGroups = 64;
Trellis2 trellis;
Trellis3 trellis;
union { __m256 vec; float val[8]; } s_helper;
union { __m256i vec; uint32_t val[8]; } o_helper;
for (int ix = 0; ix < nrc_x; ++ix) {
const float * dptr = (const float *)((const char*)vx + ix*bx);
auto d = _mm256_set1_ps(dptr[0] * 31.75f * 1.01f);
auto d = _mm256_set1_ps(dptr[0]);
auto dav = _mm256_set1_ps(dptr[1]);
const block_iq4_kt * x = (const block_iq4_kt *)(dptr + 2);
@@ -349,8 +488,8 @@ void iqk_dequantize_iq4_kt(int n, const void * vx, size_t bx, float * y, size_t
uint32_t val2 = ql[8*ib+2*j+32] + ((qh[8*ib+2*j+0] << 4) & 0xf00) + ((sh2 & 7) << 12) + o_helper.val[ib+4];
uint32_t val3 = ql[8*ib+2*j+ 1] + ((qh[8*ib+2*j+1] << 8) & 0xf00) + ((sh1 & 56) << 9) + o_helper.val[ib+0];
uint32_t val4 = ql[8*ib+2*j+33] + ((qh[8*ib+2*j+1] << 4) & 0xf00) + ((sh2 & 56) << 9) + o_helper.val[ib+4];
auto x_val1 = _mm256_fmadd_ps(scale1, trellis_gen8(trellis.next8(val1, val3)), dav);
auto x_val2 = _mm256_fmadd_ps(scale2, trellis_gen8(trellis.next8(val2, val4)), dav);
auto x_val1 = _mm256_fmadd_ps(scale1, trellis.gen8(val1, val3), dav);
auto x_val2 = _mm256_fmadd_ps(scale2, trellis.gen8(val2, val4), dav);
_mm256_storeu_ps(y + i*QK_K + 32*ib + 8*j, x_val1);
_mm256_storeu_ps(y + i*QK_K + 32*ib + 8*j + QK_K/2, x_val2);
@@ -370,7 +509,7 @@ void mul_mat_iq4_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& inf
const int nb = n/QK_K;
constexpr int kNumGroups = 64;
Trellis2 trellis;
Trellis3 trellis;
union { __m256 vec; float val[8]; } s_helper;
union { __m256i vec; uint32_t val[8]; } o_helper;
@@ -389,7 +528,7 @@ void mul_mat_iq4_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& inf
for (int ix = 0; ix < nrc_x; ++ix) {
const float * dptr = (const float *)((const char*)vx + ix*bx);
auto d = _mm256_set1_ps(dptr[0] * 31.75f * 1.01f);
auto d = _mm256_set1_ps(dptr[0]);
auto dav = dptr[1];
const block_iq4_kt * x = (const block_iq4_kt *)(dptr + 2);
@@ -413,8 +552,8 @@ void mul_mat_iq4_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& inf
uint32_t val2 = ql[8*ib+2*j+32] + ((qh[8*ib+2*j+0] << 4) & 0xf00) + ((sh2 & 7) << 12) + o_helper.val[ib+4];
uint32_t val3 = ql[8*ib+2*j+ 1] + ((qh[8*ib+2*j+1] << 8) & 0xf00) + ((sh1 & 56) << 9) + o_helper.val[ib+0];
uint32_t val4 = ql[8*ib+2*j+33] + ((qh[8*ib+2*j+1] << 4) & 0xf00) + ((sh2 & 56) << 9) + o_helper.val[ib+4];
auto x_val1 = _mm256_mul_ps(scale1, trellis_gen8(trellis.next8(val1, val3)));
auto x_val2 = _mm256_mul_ps(scale2, trellis_gen8(trellis.next8(val2, val4)));
auto x_val1 = _mm256_mul_ps(scale1, trellis.gen8(val1, val3));
auto x_val2 = _mm256_mul_ps(scale2, trellis.gen8(val2, val4));
if constexpr (nrc_y == 1) {
auto y1 = _mm256_load_ps(y[0] + i*QK_K+32*ib+8*j+ 0);
auto y2 = _mm256_load_ps(y[0] + i*QK_K+32*ib+8*j+128);
@@ -474,7 +613,7 @@ bool iqk_dequantize_ktquants(int type, int n, const void * vx, size_t bx, void *
switch (type) {
case GGML_TYPE_IQ2_KT: iqk_dequantize_iq2_kt(n, vx, bx, (float *)y, stride_y, nrc_x); break;
case GGML_TYPE_IQ3_KT: iqk_dequantize_iq3_kt(n, vx, bx, (float *)y, stride_y, nrc_x); break;
case GGML_TYPE_IQ4_KT: iqk_dequantize_iq4_kt(n, vx, bx, (float *)y, stride_y, nrc_x); break;
case GGML_TYPE_IQ4_KT: iqk_dequantize_iq4_kt_q80_r8(n, vx, bx, y, nrc_x); break;
default: return false;
}
return true;

6
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -236,9 +236,6 @@ struct MulMat {
static inline ggml_type is_dequant_better(ggml_type type, int nrc_y) {
#ifdef __AVX2__
switch (type) {
case GGML_TYPE_IQ2_KT : return nrc_y >= 32 ? GGML_TYPE_F32 : type;
case GGML_TYPE_IQ3_KT : return nrc_y >= 32 ? GGML_TYPE_F32 : type;
case GGML_TYPE_IQ4_KT : return nrc_y >= 32 ? GGML_TYPE_F32 : type;
case GGML_TYPE_IQ2_XXS: return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ2_XS : return nrc_y >= 32 ? GGML_TYPE_Q8_K_R8 : type;
case GGML_TYPE_IQ2_S : return nrc_y >= 16 ? GGML_TYPE_Q8_K_R8 : type;
@@ -267,6 +264,9 @@ struct MulMat {
case GGML_TYPE_Q6_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ4_NL : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_Q8_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
case GGML_TYPE_IQ2_KT : return nrc_y >= 32 ? GGML_TYPE_F32 : type;
case GGML_TYPE_IQ3_KT : return nrc_y >= 32 ? GGML_TYPE_F32 : type;
case GGML_TYPE_IQ4_KT : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type;
default: break;
}
#else

29
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -7408,7 +7408,7 @@ public:
constexpr static int kNg = kBlockSize/kGroupSize;
constexpr static int kNblock = kSuperBlockSize/kBlockSize;
constexpr static int kNumVal = 1 << num_bits; // i.e, 16 bits per group of 8
constexpr static float kScale = 31.75f;
constexpr static float kScale = 1.f; //31.75f;
constexpr static bool kVerbose = false;
QuantizerIQKT(int num_clusters, int num_neighbours, int offset = 4096);
@@ -7421,15 +7421,19 @@ public:
static inline void set_values(uint32_t i, float * result, float scale, int offset = 4096) {
constexpr uint32_t ka = 89226354;
constexpr uint32_t kb = 64248484;
constexpr uint32_t kmask = 0x8fff8fff;
constexpr uint32_t km32 = 0x3b603b60;
//constexpr uint32_t kmask = 0x8fff8fff;
//constexpr uint32_t km32 = 0x3b603b60;
uint32_t x = i + offset;
uint32_t s;
auto i8 = (const int8_t *)&s;
for (int k = 0; k < kGroupSize; ++k) {
x = ka*x + kb;
uint32_t s = (x & kmask) ^ km32;
float val = GGML_FP16_TO_FP32(s & 65535) + GGML_FP16_TO_FP32(s >> 16);
if constexpr (is_abs) result[k] = scale*std::abs(val);
else result[k] = scale*val;
s = x & 0x3f3f3f3f;
result[k] = scale*(i8[0] + i8[1] + i8[2] + i8[3] - 126.f);
//uint32_t s = (x & kmask) ^ km32;
//float val = GGML_FP16_TO_FP32(s & 65535) + GGML_FP16_TO_FP32(s >> 16);
//if constexpr (is_abs) result[k] = scale*std::abs(val);
//else result[k] = scale*val;
}
}
@@ -8209,7 +8213,7 @@ size_t quantize_iq2_kt(const float * src, void * dst, int64_t nrows, int64_t n_p
void dequantize_row_iq2_kt(const block_iq2_kt * x, float * y, int64_t k) {
assert(k % QuantizerIQ2KT::kSuperBlockSize == 0);
#ifdef __AVX2__
if (iqk_dequantize_ktquants(GGML_TYPE_IQ2_KT, k, x, 0, y, 0, 1)) return;
//if (iqk_dequantize_ktquants(GGML_TYPE_IQ2_KT, k, x, 0, y, 0, 1)) return;
#endif
const int nb = k / QuantizerIQ2KT::kSuperBlockSize;
const float * dptr = (const float *)x;
@@ -8560,7 +8564,10 @@ void quantize_row_iq4_kt_impl(const float * x, void * vy, int n_per_row, const f
row_av += x[j];
amax_row = std::max(amax_row, std::abs(x[j]));
}
row_av /= n_per_row;
//!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
//row_av /= n_per_row;
row_av = 0;
//!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
dptr[1] = row_av;
if (!amax_row) {
dptr[0] = 0.f;
@@ -8593,7 +8600,7 @@ void quantize_row_iq4_kt_impl(const float * x, void * vy, int n_per_row, const f
continue;
}
float best = 0;
float scale_0 = std::max(92.f, 127.f*amax/amax_row);
float scale_0 = std::max(90.f, 124.f*amax/amax_row);
for (int itry = -kNtry; itry <= kNtry; ++itry) {
quantizer1.find_best_match( amax/(8.f*itry + scale_0), xaux, weight, best_idx);
auto [dp, score_p] = quantizer1.find_best_scale(xaux, weight, best_idx);
@@ -8724,7 +8731,7 @@ size_t quantize_iq4_kt(const float * src, void * dst, int64_t nrows, int64_t n_p
void dequantize_row_iq4_kt(const block_iq4_kt * x, float * y, int64_t k) {
#ifdef __AVX2__
if (iqk_dequantize_ktquants(GGML_TYPE_IQ4_KT, k, x, 0, y, 0, 1)) return;
//if (iqk_dequantize_ktquants(GGML_TYPE_IQ4_KT, k, x, 0, y, 0, 1)) return;
#endif
using Q = QuantizerIQ4KT;
assert(k % Q::kSuperBlockSize == 0);