nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Q3_K_R4 (#134) 

* q3_k_r4: Zen4 works, but not as good as it should be

238 t/s, so sloghtly slower than q6_k_r4.

* q3_k_r4: NEON

We get PP-512(LLaMA-3.1-8B) = 106.9 t/s.
This is 1.93X faster than q3_K_S!

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow 2024-12-11 11:19:00 +01:00 committed by GitHub
parent e78e47b857
commit 4872f2f57e
10 changed files with 356 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
examples/quantize/quantize.cpp
View File

@ -36,6 +36,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "IQ3_S", LLAMA_FTYPE_MOSTLY_IQ3_S, " 3.44 bpw quantization", },
{ "IQ3_M", LLAMA_FTYPE_MOSTLY_IQ3_M, " 3.66 bpw quantization mix", },
{ "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
{ "Q3_K_R4", LLAMA_FTYPE_MOSTLY_Q3_K_R4, "Q3_K_S repacked" },
{ "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization" , },
{ "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
{ "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },

2
ggml/include/ggml.h
View File

@ -412,6 +412,7 @@ extern "C" {
GGML_TYPE_Q4_0_R4 = 202,
GGML_TYPE_Q5_0_R4 = 206,
GGML_TYPE_Q8_0_R4 = 208,
GGML_TYPE_Q3_K_R4 = 211,
GGML_TYPE_Q4_K_R4 = 212,
GGML_TYPE_Q5_K_R4 = 213,
GGML_TYPE_Q6_K_R4 = 214,
@ -481,6 +482,7 @@ extern "C" {
GGML_FTYPE_MOSTLY_Q4_0_R4 = 202, // except 1d tensors
GGML_FTYPE_MOSTLY_Q8_0_R4 = 207, // except 1d tensors
GGML_FTYPE_MOSTLY_Q5_0_R4 = 208, // except 1d tensors
GGML_FTYPE_MOSTLY_Q3_K_R4 = 211, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_K_R4 = 212, // except 1d tensors
GGML_FTYPE_MOSTLY_Q5_K_R4 = 215, // except 1d tensors
GGML_FTYPE_MOSTLY_Q6_K_R4 = 214, // except 1d tensors

9
ggml/src/ggml-common.h
View File

@ -288,6 +288,15 @@ typedef struct {
} block_q3_K;
static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
typedef struct {
ggml_half d[4]; // super-block scales
uint8_t scales_h[QK_K/16]; // scales quantized with 6 bits (high 2 bits)
uint8_t scales_l[QK_K/8]; // scales quantized with 6 bits (low 4 bits)
uint8_t qh[QK_K/2]; // quants - high bit
uint8_t qs[QK_K]; // quants - low 2 bits
} block_q3_k_r4;
static_assert(sizeof(block_q3_k_r4) == 4*sizeof(ggml_half) + QK_K/16 + QK_K/8 + QK_K/2 + QK_K, "wrong q3_k_r4 block size/padding");
// 4-bit quantization
// 8 blocks of 32 elements each
// weight is represented as x = a * q + b

1
ggml/src/ggml-quants.c
View File

@ -15202,6 +15202,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
case GGML_TYPE_Q5_0_R4: break;
case GGML_TYPE_Q6_0_R4: break;
case GGML_TYPE_Q8_0_R4: break;
case GGML_TYPE_Q3_K_R4: break;
case GGML_TYPE_Q4_K_R4: break;
case GGML_TYPE_Q5_K_R4: break;
case GGML_TYPE_Q6_K_R4: break;

24
ggml/src/ggml.c
View File

@ -875,6 +875,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.nrows = 1,
.row_meta_size = 0,
},
[GGML_TYPE_Q3_K_R4] = {
.type_name = "q3_k_r4",
.blck_size = QK_K,
.type_size = sizeof(block_q3_K),
.is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q3_k_r4,
.from_float = quantize_row_q3_k_r4,
.from_float_ref = (ggml_from_float_t) quantize_row_q3_k_r4_ref,
.vec_dot = vec_dot_q3_k_r4_q8_k,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
.row_meta_size = 0,
},
[GGML_TYPE_Q4_K] = {
.type_name = "q4_K",
.blck_size = QK_K,
@ -4058,7 +4071,8 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
case GGML_FTYPE_MOSTLY_Q8_0: wtype = GGML_TYPE_Q8_0; break;
case GGML_FTYPE_MOSTLY_Q2_K: wtype = GGML_TYPE_Q2_K; break;
case GGML_FTYPE_MOSTLY_Q3_K: wtype = GGML_TYPE_Q3_K; break;
case GGML_FTYPE_MOSTLY_Q4_K: wtype = GGML_TYPE_Q4_K; break;
case GGML_FTYPE_MOSTLY_Q3_K_R4: wtype = GGML_TYPE_Q3_K_R4; break;
case GGML_FTYPE_MOSTLY_Q4_K: wtype = GGML_TYPE_Q4_K; break;
case GGML_FTYPE_MOSTLY_Q4_K_R4: wtype = GGML_TYPE_Q4_K_R4; break;
case GGML_FTYPE_MOSTLY_Q5_K: wtype = GGML_TYPE_Q5_K; break;
case GGML_FTYPE_MOSTLY_Q5_K_R4: wtype = GGML_TYPE_Q5_K_R4; break;
@ -10591,6 +10605,7 @@ static void ggml_compute_forward_add(
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q4_K_R4:
case GGML_TYPE_Q5_K:
@ -11043,6 +11058,7 @@ static void ggml_compute_forward_add1(
case GGML_TYPE_Q8_1:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q4_K_R4:
case GGML_TYPE_Q5_K:
@ -11192,6 +11208,7 @@ static void ggml_compute_forward_acc(
case GGML_TYPE_Q8_1:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q4_K_R4:
case GGML_TYPE_Q5_K:
@ -14387,6 +14404,7 @@ static void ggml_compute_forward_out_prod(
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q4_K_R4:
case GGML_TYPE_Q5_K:
@ -14776,6 +14794,7 @@ static void ggml_compute_forward_set(
case GGML_TYPE_Q8_1:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q4_K_R4:
case GGML_TYPE_Q5_K:
@ -15059,6 +15078,7 @@ static void ggml_compute_forward_get_rows(
case GGML_TYPE_Q8_1:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q4_K_R4:
case GGML_TYPE_Q5_K:
@ -15669,6 +15689,7 @@ static void ggml_compute_forward_clamp(
case GGML_TYPE_Q8_1:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_Q4_K:
case GGML_TYPE_Q4_K_R4:
case GGML_TYPE_Q5_K:
@ -22507,6 +22528,7 @@ size_t ggml_quantize_chunk(
case GGML_TYPE_Q8_0: result = quantize_q8_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q2_K: result = quantize_q2_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q3_K: result = quantize_q3_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q3_K_R4: result = quantize_q3_k_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q4_K: result = quantize_q4_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q4_K_R4: result = quantize_q4_k_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q5_K: result = quantize_q5_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;

162
ggml/src/iqk/iqk_mul_mat.cpp
View File

@ -3439,6 +3439,85 @@ static void mul_mat_q5_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
}
#endif
template <int nrc_y>
static void mul_mat_q3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
GGML_ASSERT(nrc_x%4 == 0);
Q8<nrc_y, block_q8_K> q8(info);
auto m4 = _mm256_set1_epi8(0xf);
auto m30 = _mm256_set1_epi8(0x30);
auto m32 = _mm256_set1_epi8(32);
auto m03 = _mm256_set1_epi8(0x03);
auto m04 = _mm256_set1_epi8(0x04);
static const uint8_t k_shuff[32] = {0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15, 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15};
auto shuff = _mm256_loadu_si256((const __m256i *)k_shuff);
__m256 d4s[nrc_y];
int nbl = n / QK_K;
__m256 acc[nrc_y] = {};
__m256i qx[4];
int8_t scales[64];
for (int ix = 0; ix < nrc_x; ix += 4) {
const block_q3_k_r4 * iq3 = (const block_q3_k_r4 *)((const char *)vx + (ix+0)*bx);
for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256
auto dl = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq3[ibl].d));
auto d4 = _mm256_set_m128(dl, dl);
for (int iy = 0; iy < nrc_y; ++iy) {
d4s[iy] = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl)));
}
auto slb = _mm256_loadu_si256((const __m256i *)iq3[ibl].scales_l);
auto shbits = _mm_loadu_si128((const __m128i *)iq3[ibl].scales_h);
auto shb = MM256_SET_M128I(_mm_srli_epi16(shbits, 2), shbits);
auto scales1 = _mm256_sub_epi8(_mm256_or_si256(_mm256_and_si256(slb, m4), _mm256_and_si256(_mm256_slli_epi16(shb, 4), m30)), m32);
auto scales2 = _mm256_sub_epi8(_mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(slb, 4), m4), _mm256_and_si256(shb, m30)), m32);
_mm256_storeu_si256((__m256i *)scales+0, scales1);
_mm256_storeu_si256((__m256i *)scales+1, scales2);
{
auto t1 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm256_extracti128_si256(scales1, 0)), shuff); // blocks 0, 1, 2, 3 for each row
auto t2 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm256_extracti128_si256(scales1, 1)), shuff); // blocks 4, 5, 6, 7 for each row
auto t3 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm256_extracti128_si256(scales2, 0)), shuff); // blocks 8, 9, 10, 11 for each row
auto t4 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm256_extracti128_si256(scales2, 1)), shuff); // blocks 12, 13, 14, 15 for each row
auto s1 = MM256_SET_M128I(_mm256_extracti128_si256(t3, 0), _mm256_extracti128_si256(t1, 0)); // blocks 0, 1, 8, 9
auto s2 = MM256_SET_M128I(_mm256_extracti128_si256(t3, 1), _mm256_extracti128_si256(t1, 1)); // blocks 2, 3, 10, 11
auto s3 = MM256_SET_M128I(_mm256_extracti128_si256(t4, 0), _mm256_extracti128_si256(t2, 0)); // blocks 4, 5, 12, 13
auto s4 = MM256_SET_M128I(_mm256_extracti128_si256(t4, 1), _mm256_extracti128_si256(t2, 1)); // blocks 6, 7, 14, 15
for (int iy = 0; iy < nrc_y; ++iy) {
auto bsums = q8.load_bsums(iy, ibl);
auto sumi = _mm256_setzero_si256();
sumi = _mm256_dpwssd_epi32(sumi, s1, _mm256_shuffle_epi32(bsums, 0x00));
sumi = _mm256_dpwssd_epi32(sumi, s2, _mm256_shuffle_epi32(bsums, 0x55));
sumi = _mm256_dpwssd_epi32(sumi, s3, _mm256_shuffle_epi32(bsums, 0xaa));
sumi = _mm256_dpwssd_epi32(sumi, s4, _mm256_shuffle_epi32(bsums, 0xff));
acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4s[iy], _mm256_set1_ps(-4.f)), _mm256_cvtepi32_ps(sumi), acc[iy]);
}
}
for (int ib = 0; ib < QK_K/32; ++ib) {
auto iscales = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i *)(scales + 8*ib)));
auto scales = _mm256_cvtepi32_ps(iscales);
auto lb = _mm256_loadu_si256((const __m256i *)iq3[ibl].qs+ib);
auto hbits = _mm_loadu_si128((const __m128i *)iq3[ibl].qh+ib);
auto hb = MM256_SET_M128I(hbits, _mm_slli_epi16(hbits, 4));
qx[0] = _mm256_or_si256(_mm256_and_si256(lb, m03), _mm256_and_si256(m04, _mm256_srli_epi16(hb, 2)));
qx[1] = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(lb, 2), m03), _mm256_and_si256(m04, _mm256_srli_epi16(hb, 3)));
qx[2] = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(lb, 4), m03), _mm256_and_si256(m04, _mm256_srli_epi16(hb, 4)));
qx[3] = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(lb, 6), m03), _mm256_and_si256(m04, _mm256_srli_epi16(hb, 5)));
for (int iy = 0; iy < nrc_y; ++iy) {
auto y = _mm256_loadu_si256((const __m256i*)q8.y[iy][ibl].qs+ib);
auto sumi = _mm256_setzero_si256();
sumi = _mm256_dpbusd_epi32(sumi, qx[0], _mm256_shuffle_epi32(y, 0x00));
sumi = _mm256_dpbusd_epi32(sumi, qx[1], _mm256_shuffle_epi32(y, 0x55));
sumi = _mm256_dpbusd_epi32(sumi, qx[2], _mm256_shuffle_epi32(y, 0xaa));
sumi = _mm256_dpbusd_epi32(sumi, qx[3], _mm256_shuffle_epi32(y, 0xff));
acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, d4s[iy]), _mm256_cvtepi32_ps(sumi), acc[iy]);
}
}
}
for (int iy = 0; iy < nrc_y; ++iy) {
auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1));
acc[iy] = _mm256_setzero_ps();
info.store(ix+0, iy, sum);
}
}
}
template <int nrc_y>
static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
GGML_ASSERT(nrc_x%4 == 0);
@ -5546,6 +5625,18 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) {
mm.funcs[7] = mul_mat_iq4_xs_r4_q8_k<8>;
expected_typeB = GGML_TYPE_Q8_K32;
break;
case GGML_TYPE_Q3_K_R4:
assert (ne00 % QK_K == 0);
mm.funcs[0] = mul_mat_q3_k_r4_q8_k<1>;
mm.funcs[1] = mul_mat_q3_k_r4_q8_k<2>;
mm.funcs[2] = mul_mat_q3_k_r4_q8_k<3>;
mm.funcs[3] = mul_mat_q3_k_r4_q8_k<4>;
mm.funcs[4] = mul_mat_q3_k_r4_q8_k<5>;
mm.funcs[5] = mul_mat_q3_k_r4_q8_k<6>;
mm.funcs[6] = mul_mat_q3_k_r4_q8_k<7>;
mm.funcs[7] = mul_mat_q3_k_r4_q8_k<8>;
expected_typeB = GGML_TYPE_Q8_K;
break;
case GGML_TYPE_Q4_K_R4:
assert (ne00 % QK_K == 0);
mm.funcs[0] = mul_mat_q4_k_r4_q8_k<1>;
@ -8273,6 +8364,73 @@ IQK_ALWAYS_INLINE void prepare_q4_k_quants(const uint8x16_t& m4, const uint8x16x
qx[7] = vshrq_n_u8(bits.val[3], 4); // 28..31
}
template <int nrc_y>
void mul_mat_q3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
GGML_ASSERT(nrc_x%4 == 0);
Q8<nrc_y, block_q8_K> q8(info);
auto mf = vdupq_n_u8(0x0f);
auto m30 = vdupq_n_u8(0x30);
auto m32 = vdupq_n_s8(-32);
auto m03 = vdupq_n_u8(0x03);
auto m04 = vdupq_n_u8(0x04);
int nbl = n / QK_K;
int8x16_t qx[4];
float32x4_t acc[nrc_y] = {};
int8x16x4_t i8scales;
int16x8x4_t i16scales;
for (int ix = 0; ix < nrc_x; ix += 4) {
const block_q3_k_r4 * iq3 = (const block_q3_k_r4 *)((const char *)vx + ix*bx);
for (int ibl = 0; ibl < nbl; ++ibl) {
int32x4_t isum[nrc_y] = {};
auto d4 = vcvt_f32_f16(vld1_f16((const float16_t *)iq3[ibl].d));
auto sl = vld1q_u8_x2(iq3[ibl].scales_l);
auto sh = vld1q_u8(iq3[ibl].scales_h);
i8scales.val[0] = vaddq_s8(m32, vorrq_u8(vandq_u8(sl.val[0], mf), vandq_u8(vshlq_n_u8(sh, 4), m30)));
i8scales.val[1] = vaddq_s8(m32, vorrq_u8(vandq_u8(sl.val[1], mf), vandq_u8(vshlq_n_u8(sh, 2), m30)));
i8scales.val[2] = vaddq_s8(m32, vorrq_u8(vshrq_n_u8(sl.val[0], 4), vandq_u8(sh, m30)));
i8scales.val[3] = vaddq_s8(m32, vorrq_u8(vshrq_n_u8(sl.val[1], 4), vandq_u8(vshrq_n_u8(sh, 2), m30)));
for (int is = 0; is < 2; ++is) {
i16scales.val[0] = vmovl_s8(vget_low_s8 (i8scales.val[2*is+0]));
i16scales.val[1] = vmovl_s8(vget_high_s8(i8scales.val[2*is+0]));
i16scales.val[2] = vmovl_s8(vget_low_s8 (i8scales.val[2*is+1]));
i16scales.val[3] = vmovl_s8(vget_high_s8(i8scales.val[2*is+1]));
for (int ib = 0; ib < 4; ++ib) {
auto lbits = vld1q_u8_x2(iq3[ibl].qs + 128*is + 32*ib);
auto hbits = vld1q_u8(iq3[ibl].qh + 64*is + 16*ib);
hbits = veorq_u8(hbits, vdupq_n_u8(0xff));
auto scales = vmovl_s16(vget_low_s16 (i16scales.val[ib]));
qx[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8( lbits.val[0], m03)), vreinterpretq_s8_u8(vandq_u8(m04, vshlq_n_u8(hbits, 2))));
qx[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(lbits.val[0], 2), m03)), vreinterpretq_s8_u8(vandq_u8(m04, vshlq_n_u8(hbits, 1))));
qx[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(lbits.val[0], 4), m03)), vreinterpretq_s8_u8(vandq_u8(m04, hbits)));
qx[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(lbits.val[0], 6), m03)), vreinterpretq_s8_u8(vandq_u8(m04, vshrq_n_u8(hbits, 1))));
for (int iy = 0; iy < nrc_y; ++iy) {
auto y = vld1q_s8(q8.y[iy][ibl].qs+128*is+32*ib);
auto sumi = interleaved_dotq(qx, y);
isum[iy] = vmlaq_s32(isum[iy], scales, sumi);
}
scales = vmovl_s16(vget_high_s16(i16scales.val[ib]));
qx[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8( lbits.val[1], m03)), vreinterpretq_s8_u8(vandq_u8(m04, vshrq_n_u8(hbits, 2))));
qx[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(lbits.val[1], 2), m03)), vreinterpretq_s8_u8(vandq_u8(m04, vshrq_n_u8(hbits, 3))));
qx[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(lbits.val[1], 4), m03)), vreinterpretq_s8_u8(vandq_u8(m04, vshrq_n_u8(hbits, 4))));
qx[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(lbits.val[1], 6), m03)), vreinterpretq_s8_u8(vandq_u8(m04, vshrq_n_u8(hbits, 5))));
for (int iy = 0; iy < nrc_y; ++iy) {
auto y = vld1q_s8(q8.y[iy][ibl].qs+128*is+32*ib+16);
auto sumi = interleaved_dotq(qx, y);
isum[iy] = vmlaq_s32(isum[iy], scales, sumi);
}
}
}
for (int iy = 0; iy < nrc_y; ++iy) {
acc[iy] = vfmaq_f32(acc[iy], vmulq_f32(d4, vdupq_n_f32(q8.scale(iy, ibl))), vcvtq_f32_s32(isum[iy]));
}
}
for (int iy = 0; iy < nrc_y; ++iy) {
info.store(ix, iy, acc[iy]);
acc[iy] = vdupq_n_f32(0.f);
}
}
}
template <int nrc_y>
void mul_mat_q4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
GGML_ASSERT(nrc_x%4 == 0);
@ -8874,6 +9032,10 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) {
SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq4_xs_r4_q8_k);
expected_Btype = GGML_TYPE_Q8_K;
break;
case GGML_TYPE_Q3_K_R4:
SET_MUL_MAT_FUNCTIONS(m, mul_mat_q3_k_r4_q8_k);
expected_Btype = GGML_TYPE_Q8_K;
break;
case GGML_TYPE_Q4_K_R4:
SET_MUL_MAT_FUNCTIONS(m, mul_mat_q4_k_r4_q8_k);
expected_Btype = GGML_TYPE_Q8_K32;

136
ggml/src/iqk/iqk_quantize.cpp
View File

@ -4301,3 +4301,139 @@ void vec_dot_q5_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t b
GGML_UNUSED(by);
}
//
// ========================================= q3_k_r4
//
void quantize_row_q3_k_r4_ref(const float * x, block_q3_k_r4 * y, int64_t k) {
quantize_q3_k_r4(x, (void *)y, 4, k/4, nullptr);
}
void quantize_row_q3_k_r4(const float * x, void * y, int64_t k) {
quantize_q3_k_r4(x, y, 4, k/4, nullptr);
}
namespace {
inline void convert_q3_k(const block_q3_K& x, uint8_t * L, uint8_t * Ld) {
constexpr uint32_t kmask1 = 0x03030303;
constexpr uint32_t kmask2 = 0x0f0f0f0f;
uint32_t aux[4];
memcpy(aux, x.scales, 12);
uint32_t tmp = aux[2];
aux[2] = ((aux[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
aux[3] = ((aux[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
aux[0] = (aux[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
aux[1] = (aux[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
std::memcpy(Ld, aux, 16);
const uint8_t * q = x.qs;
const uint8_t * hm = x.hmask;
uint8_t m = 1;
for (int n = 0; n < QK_K; n += 128) {
int shift = 0;
for (int j = 0; j < 4; ++j) {
for (int l = 0; l < 32; ++l) {
*L++ = ((q[l] >> shift) & 3) + ((hm[l] & m) ? 4 : 0);
}
shift += 2;
m <<= 1;
}
q += 32;
}
}
}
static void repack_q3_k(int nrows, int n_per_row, const block_q3_K * x, block_q3_k_r4 * y) {
GGML_ASSERT(nrows%4 == 0);
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
const block_q3_K * x4[4];
uint8_t L[QK_K], Ld[QK_K/16];
for (int row = 0; row < nrows; row += 4) {
for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k;
for (int ibl = 0; ibl < nblock; ++ibl) {
std::memset(y[ibl].scales_l, 0, QK_K/8);
std::memset(y[ibl].scales_h, 0, QK_K/16);
for (int k = 0; k < 4; ++k) {
y[ibl].d[k] = x4[k][ibl].d;
convert_q3_k(x4[k][ibl], L, Ld);
for (int ib = 0; ib < QK_K/32; ++ib) {
int is = 8*ib+k;
y[ibl].scales_l[is%32] |= (Ld[2*ib+0] & 0xf) << 4*(is/32);
y[ibl].scales_h[is%16] |= (Ld[2*ib+0] >> 4) << 2*(is/16);
is += 4;
y[ibl].scales_l[is%32] |= (Ld[2*ib+1] & 0xf) << 4*(is/32);
y[ibl].scales_h[is%16] |= (Ld[2*ib+1] >> 4) << 2*(is/16);
for (int i = 0; i < 4; ++i) {
y[ibl].qs[32*ib+4*k+i+ 0] = ((L[32*ib+i+ 0] & 0x3) << 0) | ((L[32*ib+i+ 4] & 0x3) << 2) | ((L[32*ib+i+ 8] & 0x3) << 4) | ((L[32*ib+i+12] & 0x3) << 6);
y[ibl].qs[32*ib+4*k+i+16] = ((L[32*ib+i+16] & 0x3) << 0) | ((L[32*ib+i+20] & 0x3) << 2) | ((L[32*ib+i+24] & 0x3) << 4) | ((L[32*ib+i+28] & 0x3) << 6);
y[ibl].qh[16*ib+4*k+i+ 0] = ((L[32*ib+i+ 0] >> 2) << 0) | ((L[32*ib+i+ 4] >> 2) << 1) | ((L[32*ib+i+ 8] >> 2) << 2) | ((L[32*ib+i+12] >> 2) << 3)
| ((L[32*ib+i+16] >> 2) << 4) | ((L[32*ib+i+20] >> 2) << 5) | ((L[32*ib+i+24] >> 2) << 6) | ((L[32*ib+i+28] >> 2) << 7);
}
}
}
}
x += 4*nblock;
y += nblock;
}
}
size_t quantize_q3_k_r4(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) {
GGML_ASSERT(nrows%4 == 0);
GGML_ASSERT(n_per_row%QK_K == 0);
char * qcur = (char *)dst;
auto row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row);
std::vector<char> qtmp(4*row_size);
for (int row = 0; row < nrows; row += 4) {
quantize_q3_K(src, (void *)qtmp.data(), 4, n_per_row, imatrix);
repack_q3_k(4, n_per_row, (const block_q3_K *)qtmp.data(), (block_q3_k_r4 *)qcur);
qcur += 4*row_size;
src += 4*n_per_row;
}
return nrows*row_size;
}
void dequantize_row_q3_k_r4(const block_q3_k_r4 * x, float * y, int64_t k) {
auto n_per_row = k/4;
float * y4[4] = {y, y + n_per_row, y + 2*n_per_row, y + 3*n_per_row};
int nblock = n_per_row/QK_K;
for (int ibl = 0; ibl < nblock; ++ibl) {
for (int k = 0; k < 4; ++k) {
const float d = GGML_FP16_TO_FP32(x[ibl].d[k]);
auto ql = x[ibl].qs;
auto qh = x[ibl].qh;
for (int ib = 0; ib < QK_K/32; ++ib) {
int is = 8*ib + k;
float dl1 = d * ((((x[ibl].scales_l[is%32] >> 4*(is/32)) & 0xf) | (((x[ibl].scales_h[is%16] >> 2*(is/16)) & 0x03) << 4)) - 32);
is += 4;
float dl2 = d * ((((x[ibl].scales_l[is%32] >> 4*(is/32)) & 0xf) | (((x[ibl].scales_h[is%16] >> 2*(is/16)) & 0x03) << 4)) - 32);
for (int i = 0; i < 4; ++i) {
y4[k][QK_K*ibl+32*ib+i+ 0] = dl1 * ((((ql[4*k+i+ 0] >> 0) & 3) | ((qh[4*k+i] << 2) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+ 4] = dl1 * ((((ql[4*k+i+ 0] >> 2) & 3) | ((qh[4*k+i] << 1) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+ 8] = dl1 * ((((ql[4*k+i+ 0] >> 4) & 3) | ((qh[4*k+i] << 0) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+12] = dl1 * ((((ql[4*k+i+ 0] >> 6) & 3) | ((qh[4*k+i] >> 1) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+16] = dl2 * ((((ql[4*k+i+16] >> 0) & 3) | ((qh[4*k+i] >> 2) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+20] = dl2 * ((((ql[4*k+i+16] >> 2) & 3) | ((qh[4*k+i] >> 3) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+24] = dl2 * ((((ql[4*k+i+16] >> 4) & 3) | ((qh[4*k+i] >> 4) & 4)) - 4);
y4[k][QK_K*ibl+32*ib+i+28] = dl2 * ((((ql[4*k+i+16] >> 6) & 3) | ((qh[4*k+i] >> 5) & 4)) - 4);
}
ql += 32;
qh += 16;
}
}
}
}
void vec_dot_q3_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) {
#if GGML_USE_IQK_MULMAT
if (iqk_mul_mat(1, 1, n, GGML_TYPE_Q3_K_R4, vx, 0, GGML_TYPE_Q8_K, vy, 0, s, 0, 0, 1)) {
return;
}
#endif
GGML_ASSERT(n%QK4_NL == 0);
GGML_ASSERT(nrc == 1);
GGML_UNUSED(bs);
GGML_UNUSED(bx);
GGML_UNUSED(by);
}

6
ggml/src/iqk/iqk_quantize.h
View File

@ -109,6 +109,12 @@ void dequantize_row_iq2_bn_r4(const block_iq2_bn * GGML_RESTRICT x, float * GG
size_t quantize_iq2_bn_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void vec_dot_iq2_bn_r4_q8_K64(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q3_k_r4_ref(const float * GGML_RESTRICT x, block_q3_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q3_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void dequantize_row_q3_k_r4(const block_q3_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void vec_dot_q3_k_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void quantize_row_q4_k_r4_ref(const float * GGML_RESTRICT x, block_q4_k_r4 * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
size_t quantize_q4_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);

1
include/llama.h
View File

@ -183,6 +183,7 @@ extern "C" {
LLAMA_FTYPE_MOSTLY_Q4_0_R4 = 202, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q8_0_R4 = 207, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_0_R4 = 208, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q3_K_R4 = 211, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_K_R4 = 214, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_K_R4 = 216, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q6_K_R4 = 218, // except 1d tensors

18
src/llama.cpp
View File

@ -3836,6 +3836,7 @@ struct llama_model_loader {
case GGML_TYPE_Q8_0: ftype = LLAMA_FTYPE_MOSTLY_Q8_0; break;
case GGML_TYPE_Q2_K: ftype = LLAMA_FTYPE_MOSTLY_Q2_K; break;
case GGML_TYPE_Q3_K: ftype = LLAMA_FTYPE_MOSTLY_Q3_K_M; break;
case GGML_TYPE_Q3_K_R4: ftype = LLAMA_FTYPE_MOSTLY_Q3_K_R4; break;
case GGML_TYPE_Q4_K: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M; break;
case GGML_TYPE_Q4_K_R4: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_R4; break;
case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break;
@ -4548,6 +4549,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small";
case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium";
case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "Q3_K - Large";
case LLAMA_FTYPE_MOSTLY_Q3_K_R4: return "Q3_K_R4";
case LLAMA_FTYPE_MOSTLY_Q4_K_S: return "Q4_K - Small";
case LLAMA_FTYPE_MOSTLY_Q4_K_R4: return "Q4_K_R4";
case LLAMA_FTYPE_MOSTLY_Q4_K_M: return "Q4_K - Medium";
@ -15792,6 +15794,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (new_type == GGML_TYPE_IQ4_XS_R4) {
new_type = GGML_TYPE_IQ4_XS;
}
else if (new_type == GGML_TYPE_Q3_K_R4) {
new_type = GGML_TYPE_Q3_K;
}
else if (new_type == GGML_TYPE_Q4_K_R4) {
new_type = GGML_TYPE_Q4_K;
}
@ -15904,6 +15909,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (qs.model.hparams.n_gqa() >= 4) {
if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_IQ3_XXS) new_type = GGML_TYPE_IQ3_S;
else if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_IQ3_S ) new_type = GGML_TYPE_Q4_K;
else if (new_type == GGML_TYPE_Q3_K_R4) new_type = GGML_TYPE_Q4_K_R4;
else if (new_type == GGML_TYPE_Q4_K || new_type == GGML_TYPE_IQ4_XS) new_type = GGML_TYPE_Q5_K;
else if (new_type == GGML_TYPE_IQ4_NL) new_type = GGML_TYPE_Q5_K;
else if (new_type == GGML_TYPE_IQ4_NL_R4) new_type = GGML_TYPE_Q5_K;
@ -15935,7 +15941,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) {
if (qs.model.hparams.n_vocab >= 127999 && (qs.model.type == MODEL_8B || qs.model.type == MODEL_70B))
new_type = GGML_TYPE_Q4_K;
}
}
} else if (name.find("ffn_down") != std::string::npos) {
auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
int i_layer = info.first, n_layer = info.second;
@ -16003,7 +16009,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S ||
ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_K ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 ||
ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4) {
ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R4 || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_R4) {
new_type = GGML_TYPE_Q5_K;
}
} else {
@ -16073,7 +16079,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
new_type == GGML_TYPE_IQ5_K || new_type == GGML_TYPE_IQ3_K || new_type == GGML_TYPE_Q4_K_R4 ||
new_type == GGML_TYPE_IQ6_K || new_type == GGML_TYPE_IQ4_KS || new_type == GGML_TYPE_IQ4_XS_R4 ||
new_type == GGML_TYPE_IQ2_KS || new_type == GGML_TYPE_IQ4_KSS || new_type == GGML_TYPE_Q6_K_R4 ||
new_type == GGML_TYPE_Q5_K_R4) {
new_type == GGML_TYPE_Q5_K_R4 || new_type == GGML_TYPE_Q3_K_R4) {
int nx = tensor->ne[0];
int ny = tensor->ne[1];
if (nx % QK_K != 0) {
@ -16101,6 +16107,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
case GGML_TYPE_IQ1_M:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_Q3_K_R4:
case GGML_TYPE_IQ2_K:
case GGML_TYPE_IQ3_K:
case GGML_TYPE_IQ4_KSS:
@ -16201,6 +16208,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_Q3_K_S:
case LLAMA_FTYPE_MOSTLY_Q3_K_M:
case LLAMA_FTYPE_MOSTLY_Q3_K_L: default_type = GGML_TYPE_Q3_K; break;
case LLAMA_FTYPE_MOSTLY_Q3_K_R4: default_type = GGML_TYPE_Q3_K_R4; break;
case LLAMA_FTYPE_MOSTLY_Q4_K_S:
case LLAMA_FTYPE_MOSTLY_Q4_K_M: default_type = GGML_TYPE_Q4_K; break;
case LLAMA_FTYPE_MOSTLY_Q4_K_R4: default_type = GGML_TYPE_Q4_K_R4; break;
@ -16608,6 +16616,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q8_0;
else chunk_size_multiplier = 4;
}
else if (new_type == GGML_TYPE_Q3_K_R4) {
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q3_K;
else chunk_size_multiplier = 4;
}
else if (new_type == GGML_TYPE_Q4_K_R4) {
if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_K;
else chunk_size_multiplier = 4;