diff options
| -rw-r--r-- | examples/quantize/quantize.cpp | 1 | ||||
| -rw-r--r-- | ggml/include/ggml.h | 2 | ||||
| -rw-r--r-- | ggml/src/ggml-common.h | 9 | ||||
| -rw-r--r-- | ggml/src/ggml-quants.c | 1 | ||||
| -rw-r--r-- | ggml/src/ggml.c | 22 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_mul_mat.cpp | 280 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.cpp | 114 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.h | 6 | ||||
| -rw-r--r-- | include/llama.h | 5 | ||||
| -rw-r--r-- | src/llama.cpp | 19 |
10 files changed, 443 insertions, 16 deletions
diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 4c650b87..f1e3902c 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -57,6 +57,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = { { "IQ3_K", LLAMA_FTYPE_MOSTLY_IQ3_K, " 3.44 bpw non-linear quantization", }, { "IQ3_KL", LLAMA_FTYPE_MOSTLY_IQ3_KL, " 4 bpw non-linear quantization mix",}, { "IQ4_K", LLAMA_FTYPE_MOSTLY_IQ4_K, " 4.5 bpw non-linear quantization", }, + { "IQ4_K_R4", LLAMA_FTYPE_MOSTLY_IQ4_K_R4, "IQ4_K repacked", }, { "IQ5_K", LLAMA_FTYPE_MOSTLY_IQ5_K, " 5.5 bpw non-linear quantization", }, { "IQ6_K", LLAMA_FTYPE_MOSTLY_IQ6_K, " 6.6 bpw non-linear quantization", }, { "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", }, diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 2ed0fb1f..8fb4a472 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -421,6 +421,7 @@ extern "C" { GGML_TYPE_IQ4_XS_R4 = 223, GGML_TYPE_Q6_0_R4 = 233, GGML_TYPE_IQ2_BN_R4 = 335, + GGML_TYPE_IQ4_K_R4 = 339, GGML_TYPE_COUNT, }; @@ -492,6 +493,7 @@ extern "C" { GGML_FTYPE_MOSTLY_IQ4_XS_R4 = 222, // except 1d tensors GGML_FTYPE_MOSTLY_Q6_0_R4 = 227, // except 1d tensors GGML_FTYPE_MOSTLY_IQ2_BN_R4 = 329, // except 1d tensors + GGML_FTYPE_MOSTLY_IQ4_K_R4 = 332, // except 1d tensors }; // available tensor operations: diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index 61e8dfd3..2cacc711 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -542,6 +542,15 @@ typedef struct { static_assert(sizeof(block_iq4_k) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/2 + 3*QK_K/64, "wrong iq4_k block size/padding"); typedef struct { + ggml_half d[4]; + uint8_t extra[8]; + uint8_t scales_h[QK_K/16]; + uint8_t scales_l[QK_K/8]; + uint8_t qs[QK_K*2]; +} block_iq4_k_r4; +static_assert(sizeof(block_iq4_k_r4) == 4*sizeof(block_iq4_k), "wrong iq4_k_r4 block size/padding"); + +typedef struct { ggml_half d; uint16_t extra; uint8_t scales_h[QK_K/64]; diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index ff857087..64bd9459 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -15207,6 +15207,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte case GGML_TYPE_Q4_K_R4: break; case GGML_TYPE_Q5_K_R4: break; case GGML_TYPE_Q6_K_R4: break; + case GGML_TYPE_IQ4_K_R4: break; case GGML_TYPE_Q4_0_4_4: case GGML_TYPE_Q4_0_4_8: { diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 6c574933..26ca7991 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1313,6 +1313,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .nrows = 1, .row_meta_size = 0, }, + [GGML_TYPE_IQ4_K_R4] = { + .type_name = "iq4_k_r4", + .blck_size = QK_K, + .type_size = sizeof(block_iq4_k), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_iq4_k_r4, + .from_float = quantize_row_iq4_k_r4, + .from_float_ref = (ggml_from_float_t)quantize_row_iq4_k_r4_ref, + .vec_dot = vec_dot_iq4_k_r4_q8_k, + .vec_dot_type = GGML_TYPE_Q8_K, + .nrows = 1, + .row_meta_size = 0, + }, [GGML_TYPE_IQ5_K] = { .type_name = "iq5_k", .blck_size = QK_K, @@ -4114,6 +4127,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_IQ2_KS: wtype = GGML_TYPE_IQ2_KS; break; case GGML_FTYPE_MOSTLY_IQ3_K: wtype = GGML_TYPE_IQ3_K; break; case GGML_FTYPE_MOSTLY_IQ4_K: wtype = GGML_TYPE_IQ4_K; break; + case GGML_FTYPE_MOSTLY_IQ4_K_R4: wtype = GGML_TYPE_IQ4_K_R4; break; case GGML_FTYPE_MOSTLY_IQ5_K: wtype = GGML_TYPE_IQ5_K; break; case GGML_FTYPE_MOSTLY_IQ6_K: wtype = GGML_TYPE_IQ6_K; break; case GGML_FTYPE_MOSTLY_IQ3_S: wtype = GGML_TYPE_IQ3_S; break; @@ -10649,6 +10663,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: + case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ5_K: case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: @@ -11103,6 +11118,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: + case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ5_K: case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: @@ -11254,6 +11270,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: + case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ5_K: case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: @@ -14451,6 +14468,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: + case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ5_K: case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: @@ -14842,6 +14860,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: + case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ5_K: case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: @@ -15127,6 +15146,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: + case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ5_K: case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: @@ -15739,6 +15759,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: + case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ5_K: case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: @@ -22581,6 +22602,7 @@ size_t ggml_quantize_chunk( case GGML_TYPE_IQ2_KS: result = quantize_iq2_ks (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ3_K: result = quantize_iq3_k (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_K: result = quantize_iq4_k (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ4_K_R4:result = quantize_iq4_k_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ5_K: result = quantize_iq5_k (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ6_K: result = quantize_iq6_k (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp index f0e9d61d..6bfd4f77 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -3785,6 +3785,125 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn } } +template <int nrc_y> +static void mul_mat_iq4_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 ms = _mm256_set1_epi8(4); + //auto shift_shuffle = _mm256_set_epi64x(0x0303030302020202, 0x0101010100000000, 0x0303030302020202, 0x0101010100000000); + auto shift_shuffle = _mm256_set_epi64x(0x0707070706060606, 0x0505050504040404, 0x0303030302020202, 0x0101010100000000); +#ifdef HAVE_FANCY_SIMD + auto values = load_iq4nl_values_256(); + __m256 d4s[nrc_y]; + 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); +#else + auto m1 = _mm256_set1_epi16(1); + auto values128 = _mm_loadu_si128((const __m128i *)iq4k_values); + auto values = MM256_SET_M128I(values128, values128); +#endif + int nbl = n / QK_K; + __m256 acc[nrc_y] = {}; + __m256i qx[4]; + int8_t stored_scales[64]; + for (int ix = 0; ix < nrc_x; ix += 4) { + const block_iq4_k_r4 * iq4 = (const block_iq4_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 *)iq4[ibl].d)); + auto d4 = _mm256_set_m128(dl, dl); + auto extra = _mm256_set1_epi64x(*(const uint64_t *)iq4[ibl].extra); +#ifdef HAVE_FANCY_SIMD + for (int iy = 0; iy < nrc_y; ++iy) { + d4s[iy] = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl))); + } +#else + if constexpr (nrc_y == 1) { + d4 = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(0, ibl))); + } +#endif + auto slbits = _mm256_loadu_si256((const __m256i *)iq4[ibl].scales_l); + auto sl1 = _mm256_and_si256(slbits, m4); + auto sl2 = _mm256_and_si256(_mm256_srli_epi16(slbits, 4), m4); + auto shbits = _mm_loadu_si128((const __m128i*)iq4[ibl].scales_h); + auto sh = MM256_SET_M128I(_mm_srli_epi16(shbits, 2), shbits); + auto i8scales1 = _mm256_sub_epi8(_mm256_or_si256(sl1, _mm256_and_si256(m30, _mm256_slli_epi16(sh, 4))), m32); + auto i8scales2 = _mm256_sub_epi8(_mm256_or_si256(sl2, _mm256_and_si256(m30, sh)), m32); + _mm256_storeu_si256((__m256i *)stored_scales+0, i8scales1); + _mm256_storeu_si256((__m256i *)stored_scales+1, i8scales2); +#ifdef HAVE_FANCY_SIMD + { + auto t1 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm256_extracti128_si256(i8scales1, 0)), shuff); // blocks 0, 1, 2, 3 for each row + auto t2 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm256_extracti128_si256(i8scales1, 1)), shuff); // blocks 4, 5, 6, 7 for each row + auto t3 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm256_extracti128_si256(i8scales2, 0)), shuff); // blocks 8, 9, 10, 11 for each row + auto t4 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm256_extracti128_si256(i8scales2, 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(-128.f)), _mm256_cvtepi32_ps(sumi), acc[iy]); + } + } +#endif + for (int ib = 0; ib < QK_K/32; ++ib) { + auto iscales = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i *)(stored_scales + 8*ib))); +#ifdef HAVE_FANCY_SIMD + auto scales = _mm256_cvtepi32_ps(iscales); +#else + auto scales = _mm256_mul_ps(d4, _mm256_cvtepi32_ps(iscales)); +#endif + auto bits1 = _mm256_loadu_si256((const __m256i *)iq4[ibl].qs+2*ib+0); + auto bits2 = _mm256_loadu_si256((const __m256i *)iq4[ibl].qs+2*ib+1); + auto shift = _mm256_and_si256(ms, _mm256_slli_epi16(extra, 2)); extra = _mm256_srli_epi16(extra, 1); + shift = _mm256_shuffle_epi8(shift, shift_shuffle); + qx[0] = _mm256_add_epi8(shift, _mm256_shuffle_epi8(values, _mm256_and_si256(bits1, m4))); + qx[1] = _mm256_add_epi8(shift, _mm256_shuffle_epi8(values, _mm256_and_si256(bits2, m4))); + qx[2] = _mm256_add_epi8(shift, _mm256_shuffle_epi8(values, _mm256_and_si256(_mm256_srli_epi16(bits1, 4), m4))); + qx[3] = _mm256_add_epi8(shift, _mm256_shuffle_epi8(values, _mm256_and_si256(_mm256_srli_epi16(bits2, 4), m4))); +#ifndef HAVE_FANCY_SIMD + auto s1 = _mm256_sign_epi8(qx[0], qx[0]); + auto s2 = _mm256_sign_epi8(qx[1], qx[1]); + auto s3 = _mm256_sign_epi8(qx[2], qx[2]); + auto s4 = _mm256_sign_epi8(qx[3], qx[3]); +#endif + for (int iy = 0; iy < nrc_y; ++iy) { + auto y = _mm256_loadu_si256((const __m256i*)q8.y[iy][ibl].qs+ib); +#ifdef HAVE_FANCY_SIMD + 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]); +#else + auto sumi1 = _mm256_maddubs_epi16(s1, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0x00), qx[0])); + auto sumi2 = _mm256_maddubs_epi16(s2, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0x55), qx[1])); + auto sumi3 = _mm256_maddubs_epi16(s3, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0xaa), qx[2])); + auto sumi4 = _mm256_maddubs_epi16(s4, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0xff), qx[3])); + auto sumi = _mm256_add_epi32(_mm256_add_epi32(_mm256_madd_epi16(m1, sumi1), _mm256_madd_epi16(m1, sumi2)), + _mm256_add_epi32(_mm256_madd_epi16(m1, sumi3), _mm256_madd_epi16(m1, sumi4))); + acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(sumi), acc[iy]); +#endif + } + } + } + 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 <typename Bits> inline void multiply_add_1(int j, const Bits& bits, const __m256i * scales, const __m256i * q8, __m256i * sumi) { if (j == 0) { @@ -5804,18 +5923,6 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) { 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>; - mm.funcs[1] = mul_mat_q4_k_r4_q8_k<2>; - mm.funcs[2] = mul_mat_q4_k_r4_q8_k<3>; - mm.funcs[3] = mul_mat_q4_k_r4_q8_k<4>; - mm.funcs[4] = mul_mat_q4_k_r4_q8_k<5>; - mm.funcs[5] = mul_mat_q4_k_r4_q8_k<6>; - mm.funcs[6] = mul_mat_q4_k_r4_q8_k<7>; - mm.funcs[7] = mul_mat_q4_k_r4_q8_k<8>; - expected_typeB = GGML_TYPE_Q8_K32; - break; case GGML_TYPE_Q5_K_R4: assert (ne00 % QK_K == 0); mm.funcs[0] = mul_mat_q5_k_r4_q8_k<1>; @@ -5840,6 +5947,18 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) { mm.funcs[7] = mul_mat_q6_k_r4_q8_k<8>; expected_typeB = GGML_TYPE_Q8_K; break; + case GGML_TYPE_IQ4_K_R4: + assert (ne00 % QK_K == 0); + mm.funcs[0] = mul_mat_iq4_k_r4_q8_k<1>; + mm.funcs[1] = mul_mat_iq4_k_r4_q8_k<2>; + mm.funcs[2] = mul_mat_iq4_k_r4_q8_k<3>; + mm.funcs[3] = mul_mat_iq4_k_r4_q8_k<4>; + mm.funcs[4] = mul_mat_iq4_k_r4_q8_k<5>; + mm.funcs[5] = mul_mat_iq4_k_r4_q8_k<6>; + mm.funcs[6] = mul_mat_iq4_k_r4_q8_k<7>; + mm.funcs[7] = mul_mat_iq4_k_r4_q8_k<8>; + expected_typeB = GGML_TYPE_Q8_K; + break; case GGML_TYPE_Q4_0_R4: assert (ne00 % QK4_NL == 0); mm.funcs[0] = mul_mat_q4_0_r4_q8_1<1>; @@ -8516,6 +8635,139 @@ void mul_mat_iq4_xs_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& i } } +template <int nrc_y> +void mul_mat_iq4_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 = vdupq_n_u8(0xf); + auto m3 = vdupq_n_u8(0x30); + auto ms = vdupq_n_u8(4); + auto m32 = vdupq_n_s8(-32); + uint8x16x2_t shift_shuffle = { + vreinterpretq_u8_u64(uint64x2_t{0x0101010100000000, 0x0303030302020202}), + vreinterpretq_u8_u64(uint64x2_t{0x0505050504040404, 0x0707070706060606}) + }; + auto values = vld1q_s8(iq4k_values); + int nbl = n / QK_K; + int8x16_t qx[4]; + int8x16x4_t i8scales; + int16x8x4_t i16scales; + float32x4_t acc[nrc_y] = {}; + for (int ix = 0; ix < nrc_x; ix += 4) { + const block_iq4_k_r4 * iq4 = (const block_iq4_k_r4 *)((const char *)vx + ix*bx); + for (int ibl = 0; ibl < nbl; ++ibl) { + auto d4 = vcvt_f32_f16(vld1_f16((const float16_t *)iq4[ibl].d)); + auto extra8 = vld1_u8(iq4[ibl].extra); + uint8x16_t extra; + if constexpr (nrc_y == 1) { + extra = vcombine_u8(extra8, vshr_n_u8(extra8,1)); + } else { + extra = vcombine_u8(extra8, extra8); + } + auto sl = vld1q_u8_x2(iq4[ibl].scales_l); + auto sh = vld1q_u8(iq4[ibl].scales_h); + i8scales.val[0] = vaddq_s8(vorrq_u8(vandq_u8(sl.val[0], m4), vandq_u8(vshlq_n_u8(sh, 4), m3)), m32); + i8scales.val[1] = vaddq_s8(vorrq_u8(vandq_u8(sl.val[1], m4), vandq_u8(vshlq_n_u8(sh, 2), m3)), m32); + i8scales.val[2] = vaddq_s8(vorrq_u8(vshrq_n_u8(sl.val[0], 4), vandq_u8(sh, m3)), m32); + i8scales.val[3] = vaddq_s8(vorrq_u8(vshrq_n_u8(sl.val[1], 4), vandq_u8(vshrq_n_u8(sh, 2), m3)), m32); + int32x4_t isum[nrc_y] = {}; + if constexpr (nrc_y == 1) { + auto s8_1 = vmulq_s8(i8scales.val[0], vandq_u8(ms, vshlq_n_u8(extra, 2))); + auto s8_2 = vmulq_s8(i8scales.val[1], vandq_u8(ms, extra)); + auto s16_1 = vmovl_s8(vget_low_s8 (s8_1)); + auto s16_2 = vmovl_s8(vget_high_s8(s8_1)); + auto s16_3 = vmovl_s8(vget_low_s8 (s8_2)); + auto s16_4 = vmovl_s8(vget_high_s8(s8_2)); + for (int iy = 0; iy < nrc_y; ++iy) { + auto b8 = vld1_s16(q8.y[iy][ibl].bsums); + isum[iy] = vmlal_lane_s16(isum[iy], vget_low_s16 (s16_1), b8, 0); + isum[iy] = vmlal_lane_s16(isum[iy], vget_high_s16(s16_1), b8, 1); + isum[iy] = vmlal_lane_s16(isum[iy], vget_low_s16 (s16_2), b8, 2); + isum[iy] = vmlal_lane_s16(isum[iy], vget_high_s16(s16_2), b8, 3); + b8 = vld1_s16(q8.y[iy][ibl].bsums+4); + isum[iy] = vmlal_lane_s16(isum[iy], vget_low_s16 (s16_3), b8, 0); + isum[iy] = vmlal_lane_s16(isum[iy], vget_high_s16(s16_3), b8, 1); + isum[iy] = vmlal_lane_s16(isum[iy], vget_low_s16 (s16_4), b8, 2); + isum[iy] = vmlal_lane_s16(isum[iy], vget_high_s16(s16_4), b8, 3); + } + s8_1 = vmulq_s8(i8scales.val[2], vandq_u8(ms, vshrq_n_u8(extra, 2))); + s8_2 = vmulq_s8(i8scales.val[3], vandq_u8(ms, vshrq_n_u8(extra, 4))); + s16_1 = vmovl_s8(vget_low_s8 (s8_1)); + s16_2 = vmovl_s8(vget_high_s8(s8_1)); + s16_3 = vmovl_s8(vget_low_s8 (s8_2)); + s16_4 = vmovl_s8(vget_high_s8(s8_2)); + for (int iy = 0; iy < nrc_y; ++iy) { + auto b8 = vld1_s16(q8.y[iy][ibl].bsums+8); + isum[iy] = vmlal_lane_s16(isum[iy], vget_low_s16 (s16_1), b8, 0); + isum[iy] = vmlal_lane_s16(isum[iy], vget_high_s16(s16_1), b8, 1); + isum[iy] = vmlal_lane_s16(isum[iy], vget_low_s16 (s16_2), b8, 2); + isum[iy] = vmlal_lane_s16(isum[iy], vget_high_s16(s16_2), b8, 3); + b8 = vld1_s16(q8.y[iy][ibl].bsums+12); + isum[iy] = vmlal_lane_s16(isum[iy], vget_low_s16 (s16_3), b8, 0); + isum[iy] = vmlal_lane_s16(isum[iy], vget_high_s16(s16_3), b8, 1); + isum[iy] = vmlal_lane_s16(isum[iy], vget_low_s16 (s16_4), b8, 2); + isum[iy] = vmlal_lane_s16(isum[iy], vget_high_s16(s16_4), b8, 3); + } + } + 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 bits = vld1q_u8_x4(iq4[ibl].qs + 256*is + 64*ib); + uint8x16_t shifts; + if constexpr (nrc_y == 1) { + qx[0] = vqtbl1q_s8(values, vandq_u8(bits.val[0], m4)); // 0...3 from the 4 rows + qx[1] = vqtbl1q_s8(values, vandq_u8(bits.val[2], m4)); // 4...7 + qx[2] = vqtbl1q_s8(values, vshrq_n_u8(bits.val[0], 4)); // 8..11 + qx[3] = vqtbl1q_s8(values, vshrq_n_u8(bits.val[2], 4)); // 12..15 + } else { + shifts = vandq_u8(ms, vshlq_n_u8(extra, 2)); + auto shift = vqtbl1q_u8(shifts, shift_shuffle.val[0]); + extra = vshrq_n_u8(extra, 1); + qx[0] = vaddq_s8(shift, vqtbl1q_s8(values, vandq_u8(bits.val[0], m4))); // 0...3 from the 4 rows + qx[1] = vaddq_s8(shift, vqtbl1q_s8(values, vandq_u8(bits.val[2], m4))); // 4...7 + qx[2] = vaddq_s8(shift, vqtbl1q_s8(values, vshrq_n_u8(bits.val[0], 4))); // 8..11 + qx[3] = vaddq_s8(shift, vqtbl1q_s8(values, vshrq_n_u8(bits.val[2], 4))); // 12..15 + } + auto scales = vmovl_s16(vget_low_s16 (i16scales.val[ib])); + 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); + } + if constexpr (nrc_y == 1) { + qx[0] = vqtbl1q_s8(values, vandq_u8(bits.val[1], m4)); // 16..19 + qx[1] = vqtbl1q_s8(values, vandq_u8(bits.val[3], m4)); // 20..23 + qx[2] = vqtbl1q_s8(values, vshrq_n_u8(bits.val[1], 4)); // 24..27 + qx[3] = vqtbl1q_s8(values, vshrq_n_u8(bits.val[3], 4)); // 28..31 + } else { + auto shift = vqtbl1q_u8(shifts, shift_shuffle.val[1]); + qx[0] = vaddq_s8(shift, vqtbl1q_s8(values, vandq_u8(bits.val[1], m4))); // 16..19 + qx[1] = vaddq_s8(shift, vqtbl1q_s8(values, vandq_u8(bits.val[3], m4))); // 20..23 + qx[2] = vaddq_s8(shift, vqtbl1q_s8(values, vshrq_n_u8(bits.val[1], 4))); // 24..27 + qx[3] = vaddq_s8(shift, vqtbl1q_s8(values, vshrq_n_u8(bits.val[3], 4))); // 28..31 + } + scales = vmovl_s16(vget_high_s16(i16scales.val[ib])); + 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); + } + } +} + IQK_ALWAYS_INLINE void prepare_q4_k_quants(const uint8x16_t& m4, const uint8x16x4_t& bits, int8x16_t * qx) { qx[0] = vandq_u8(bits.val[0], m4); // 0...3 from the 4 rows qx[1] = vandq_u8(bits.val[1], m4); // 16..19 @@ -9294,6 +9546,10 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) { SET_MUL_MAT_FUNCTIONS(m, mul_mat_q6_k_r4_q8_k); expected_Btype = GGML_TYPE_Q8_K; break; + case GGML_TYPE_IQ4_K_R4: + SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq4_k_r4_q8_k); + expected_Btype = GGML_TYPE_Q8_K; + break; case GGML_TYPE_Q4_0_R4: SET_MUL_MAT_FUNCTIONS_T(m, mul_mat_qx_r4_q8_0, Q4_0_R4_Dequantizer); expected_Btype = GGML_TYPE_Q8_0; diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp index 49e2cf8e..438a277e 100644 --- a/ggml/src/iqk/iqk_quantize.cpp +++ b/ggml/src/iqk/iqk_quantize.cpp @@ -4552,3 +4552,117 @@ void vec_dot_q2_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t b GGML_UNUSED(by); } +// +// ========================================= iq4_k_r4 +// + +void quantize_row_iq4_k_r4_ref(const float * x, block_iq4_k_r4 * y, int64_t k) { + quantize_iq4_k_r4(x, (void *)y, 4, k/4, nullptr); +} + +void quantize_row_iq4_k_r4(const float * x, void * y, int64_t k) { + quantize_iq4_k_r4(x, y, 4, k/4, nullptr); +} + +static void repack_iq4_k(int nrows, int n_per_row, const block_iq4_k * x, block_iq4_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_iq4_k * x4[4]; + 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].extra, 0, 8); + 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; + auto extra = x4[k][ibl].extra; + for (int ib = 0; ib < QK_K/32; ++ib) { + if (extra & 1) y[ibl].extra[k+0] |= (1 << ib); + if (extra & 2) y[ibl].extra[k+4] |= (1 << ib); + extra >>= 2; + uint8_t sl1 = x4[k][ibl].scales_l[ib] & 0xf; + uint8_t sl2 = x4[k][ibl].scales_l[ib] >> 4; + uint8_t sh = x4[k][ibl].scales_h[ib/2] >> 4*(ib%2); + uint8_t sh1 = (sh >> 0) & 3; + uint8_t sh2 = (sh >> 2) & 3; + int i = 8*ib + k; + y[ibl].scales_l[i%32] |= (sl1 << 4*(i/32)); + y[ibl].scales_h[i%16] |= (sh1 << 2*(i/16)); + i += 4; + y[ibl].scales_l[i%32] |= (sl2 << 4*(i/32)); + y[ibl].scales_h[i%16] |= (sh2 << 2*(i/16)); + } + } + for (int ib = 0; ib < QK_K/32; ++ib) { + for (int k = 0; k < 4; ++k) for (int i = 0; i < 4; ++i) { + y[ibl].qs[64*ib+4*k+i+ 0] = (x4[k][ibl].qs[16*ib+i+0] & 0xf) | ((x4[k][ibl].qs[16*ib+i+ 8] & 0x0f) << 4); // 0....3 + 8...11 from each row + y[ibl].qs[64*ib+4*k+i+16] = (x4[k][ibl].qs[16*ib+i+0] >> 4) | ((x4[k][ibl].qs[16*ib+i+ 8] & 0xf0)); // 16...19 + 24...27 from each row + y[ibl].qs[64*ib+4*k+i+32] = (x4[k][ibl].qs[16*ib+i+4] & 0xf) | ((x4[k][ibl].qs[16*ib+i+12] & 0x0f) << 4); // 4....7 + 12...15 from each row + y[ibl].qs[64*ib+4*k+i+48] = (x4[k][ibl].qs[16*ib+i+4] >> 4) | ((x4[k][ibl].qs[16*ib+i+12] & 0xf0)); // 20...23 + 28...31 from each row + } + } + } + x += 4*nblock; + y += nblock; + } +} + +size_t quantize_iq4_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_IQ4_K, n_per_row); + std::vector<char> qtmp(4*row_size); + for (int row = 0; row < nrows; row += 4) { + quantize_iq4_k(src, (void *)qtmp.data(), 4, n_per_row, imatrix); + repack_iq4_k(4, n_per_row, (const block_iq4_k *)qtmp.data(), (block_iq4_k_r4 *)qcur); + qcur += 4*row_size; + src += 4*n_per_row; + } + return nrows*row_size; +} + +void dequantize_row_iq4_k_r4(const block_iq4_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]); + 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)) & 3) << 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)) & 3) << 4)) - 32); + auto values1 = iq4k_values + (x[ibl].extra[k+0] & (1 << ib) ? 16 : 0); + auto values2 = iq4k_values + (x[ibl].extra[k+4] & (1 << ib) ? 16 : 0); + for (int i = 0; i < 4; ++i) { + y4[k][QK_K*ibl+32*ib+i+ 0] = dl1 * values1[x[ibl].qs[64*ib+4*k+i+ 0] & 0xf]; + y4[k][QK_K*ibl+32*ib+i+ 8] = dl1 * values1[x[ibl].qs[64*ib+4*k+i+ 0] >> 4]; + y4[k][QK_K*ibl+32*ib+i+16] = dl2 * values2[x[ibl].qs[64*ib+4*k+i+16] & 0xf]; + y4[k][QK_K*ibl+32*ib+i+24] = dl2 * values2[x[ibl].qs[64*ib+4*k+i+16] >> 4]; + y4[k][QK_K*ibl+32*ib+i+ 4] = dl1 * values1[x[ibl].qs[64*ib+4*k+i+32] & 0xf]; + y4[k][QK_K*ibl+32*ib+i+12] = dl1 * values1[x[ibl].qs[64*ib+4*k+i+32] >> 4]; + y4[k][QK_K*ibl+32*ib+i+20] = dl2 * values2[x[ibl].qs[64*ib+4*k+i+48] & 0xf]; + y4[k][QK_K*ibl+32*ib+i+28] = dl2 * values2[x[ibl].qs[64*ib+4*k+i+48] >> 4]; + } + } + } + } +} + +void vec_dot_iq4_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_IQ4_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); +} + diff --git a/ggml/src/iqk/iqk_quantize.h b/ggml/src/iqk/iqk_quantize.h index 4a1c31f8..c5702d73 100644 --- a/ggml/src/iqk/iqk_quantize.h +++ b/ggml/src/iqk/iqk_quantize.h @@ -139,6 +139,12 @@ size_t quantize_q6_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT ds void dequantize_row_q6_k_r4(const block_q6_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); void vec_dot_q6_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_iq4_k_r4_ref(const float * GGML_RESTRICT x, block_iq4_k_r4 * GGML_RESTRICT y, int64_t k); +void quantize_row_iq4_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +size_t quantize_iq4_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_iq4_k_r4(const block_iq4_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void vec_dot_iq4_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 iqk_quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); void quantize_row_q8_K64_ref(const float * GGML_RESTRICT x, block_q8_K64 * GGML_RESTRICT y, int64_t k); void quantize_row_q8_K64(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); diff --git a/include/llama.h b/include/llama.h index 0992b10a..117f1659 100644 --- a/include/llama.h +++ b/include/llama.h @@ -190,8 +190,9 @@ extern "C" { LLAMA_FTYPE_MOSTLY_Q6_K_R4 = 218, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 = 225, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ4_XS_R4 = 230, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q6_0_R4 = 235, // except 1d tensors - LLAMA_FTYPE_MOSTLY_IQ2_BN_R4 = 237, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q6_0_R4 = 335, // except 1d tensors + LLAMA_FTYPE_MOSTLY_IQ2_BN_R4 = 337, // except 1d tensors + LLAMA_FTYPE_MOSTLY_IQ4_K_R4 = 340, // except 1d tensors LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file }; diff --git a/src/llama.cpp b/src/llama.cpp index 6ecf0452..9356c639 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3866,6 +3866,7 @@ struct llama_model_loader { case GGML_TYPE_IQ2_K: ftype = LLAMA_FTYPE_MOSTLY_IQ2_K; break; case GGML_TYPE_IQ3_K: ftype = LLAMA_FTYPE_MOSTLY_IQ3_K; break; case GGML_TYPE_IQ4_K: ftype = LLAMA_FTYPE_MOSTLY_IQ4_K; break; + case GGML_TYPE_IQ4_K_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ4_K_R4;break; case GGML_TYPE_IQ5_K: ftype = LLAMA_FTYPE_MOSTLY_IQ5_K; break; case GGML_TYPE_IQ6_K: ftype = LLAMA_FTYPE_MOSTLY_IQ6_K; break; case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break; @@ -4582,6 +4583,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ3_K: return "IQ3_K - 3.4325 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_KL: return "IQ3_KL - 4 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_K: return "IQ4_K - 4.5 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ4_K_R4: return "IQ4_K_R4 - 4.5 bpw"; case LLAMA_FTYPE_MOSTLY_IQ5_K: return "IQ5_K - 5.5 bpw"; case LLAMA_FTYPE_MOSTLY_IQ6_K: return "IQ6_K - 6.6 bpw"; case LLAMA_FTYPE_MOSTLY_IQ1_BN: return "IQ1_BN - 1.625 bpw Bitnet"; @@ -15810,6 +15812,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (new_type == GGML_TYPE_Q6_K_R4) { new_type = GGML_TYPE_Q6_K; } + else if (new_type == GGML_TYPE_IQ4_K_R4) { + new_type = GGML_TYPE_IQ4_K; + } else if (new_type == GGML_TYPE_Q4_0_R4) { new_type = GGML_TYPE_Q4_0; } @@ -15894,6 +15899,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_K && qs.model.hparams.n_gqa() >= 2) { new_type = GGML_TYPE_IQ5_K; } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_K_R4 && qs.model.hparams.n_gqa() >= 2) { + new_type = GGML_TYPE_IQ5_K; + } else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) && use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; @@ -16020,7 +16028,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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_Q3_K_R4 || - ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4) { + ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4|| ftype == LLAMA_FTYPE_MOSTLY_IQ4_K_R4) { new_type = GGML_TYPE_Q5_K; } } else { @@ -16090,7 +16098,8 @@ 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_Q3_K_R4 || new_type == GGML_TYPE_Q2_K_R4) { + new_type == GGML_TYPE_Q5_K_R4 || new_type == GGML_TYPE_Q3_K_R4 || new_type == GGML_TYPE_Q2_K_R4 || + new_type == GGML_TYPE_IQ4_K_R4) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; if (nx % QK_K != 0) { @@ -16127,6 +16136,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n case GGML_TYPE_IQ4_XS_R4: case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break; case GGML_TYPE_IQ4_K: + case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break; case GGML_TYPE_IQ5_K: @@ -16255,6 +16265,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ3_K: default_type = GGML_TYPE_IQ3_K; break; case LLAMA_FTYPE_MOSTLY_IQ3_KL: default_type = GGML_TYPE_IQ3_K; break; case LLAMA_FTYPE_MOSTLY_IQ4_K: default_type = GGML_TYPE_IQ4_K; break; + case LLAMA_FTYPE_MOSTLY_IQ4_K_R4:default_type = GGML_TYPE_IQ4_K_R4;break; case LLAMA_FTYPE_MOSTLY_IQ5_K: default_type = GGML_TYPE_IQ5_K; break; case LLAMA_FTYPE_MOSTLY_IQ6_K: default_type = GGML_TYPE_IQ6_K; break; case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break; @@ -16653,6 +16664,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_BN; else chunk_size_multiplier = 4; } + else if (new_type == GGML_TYPE_IQ4_K_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ4_K; + else chunk_size_multiplier = 4; + } LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type)); fflush(stdout); |