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 | 8 | ||||
| -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 | 254 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.cpp | 123 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.h | 6 | ||||
| -rw-r--r-- | include/llama.h | 1 | ||||
| -rw-r--r-- | src/llama.cpp | 23 |
10 files changed, 430 insertions, 11 deletions
diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index a2f83150..73485838 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -53,6 +53,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = { { "IQ4_KS", LLAMA_FTYPE_MOSTLY_IQ4_KS, " 4.25 bpw non-linear quantization", }, { "IQ4_KSS", LLAMA_FTYPE_MOSTLY_IQ4_KSS, " 4.0 bpw non-linear quantization", }, { "IQ2_K", LLAMA_FTYPE_MOSTLY_IQ2_K, " 2.375 bpw non-linear quantization",}, + { "IQ2_K_R4", LLAMA_FTYPE_MOSTLY_IQ2_K_R4, "IQ2_K repacked",}, { "IQ2_KS", LLAMA_FTYPE_MOSTLY_IQ2_KS, " 2.1875 bpw non-linear quantization",}, { "IQ3_K", LLAMA_FTYPE_MOSTLY_IQ3_K, " 3.44 bpw non-linear quantization", }, { "IQ3_K_R4", LLAMA_FTYPE_MOSTLY_IQ3_K_R4, "IQ3_K repacked", }, diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 04831c2f..77ee0fb9 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -423,6 +423,7 @@ extern "C" { GGML_TYPE_BF16_R16 = 230, GGML_TYPE_Q6_0_R4 = 233, GGML_TYPE_IQ2_BN_R4 = 335, + GGML_TYPE_IQ2_K_R4 = 337, GGML_TYPE_IQ3_K_R4 = 338, GGML_TYPE_IQ4_K_R4 = 339, GGML_TYPE_Q8_K_R8 = 399, @@ -498,6 +499,7 @@ extern "C" { GGML_FTYPE_MOSTLY_BF16_R16 = 224, // 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_IQ2_K_R4 = 330, // except 1d tensors GGML_FTYPE_MOSTLY_IQ3_K_R4 = 331, // except 1d tensors GGML_FTYPE_MOSTLY_IQ4_K_R4 = 332, // except 1d tensors GGML_FTYPE_MOSTLY_Q8_K_R8 = 399, // except 1d tensors diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index ca56704c..03cc3460 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -522,6 +522,14 @@ typedef struct { static_assert(sizeof(block_iq2_k) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/32 + QK_K/4, "wrong iq2_k block size/padding"); typedef struct { + ggml_half d[4]; + uint8_t extra[8]; + uint8_t scales[QK_K/8]; + uint8_t qs[QK_K]; +} block_iq2_k_r4; +static_assert(sizeof(block_iq2_k_r4) == 4*sizeof(block_iq2_k), "wrong iq2_k_r4 block size/padding"); + +typedef struct { uint16_t extra; uint8_t scales[QK_K/64]; uint8_t qs[QK_K/4]; diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index 1d022672..a3beba20 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_IQ2_K_R4: break; case GGML_TYPE_IQ3_K_R4: break; case GGML_TYPE_IQ4_K_R4: break; case GGML_TYPE_Q8_K_R8: break; diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 4194d943..45c873f2 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1308,6 +1308,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .nrows = 1, .row_meta_size = 0, }, + [GGML_TYPE_IQ2_K_R4] = { + .type_name = "iq2_k_r4", + .blck_size = QK_K, + .type_size = sizeof(block_iq2_k), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_iq2_k_r4, + .from_float = quantize_row_iq2_k_r4, + .from_float_ref = (ggml_from_float_t)quantize_row_iq2_k_r4_ref, + .vec_dot = vec_dot_iq2_k_r4_q8_k, + .vec_dot_type = GGML_TYPE_Q8_K, + .nrows = 1, + .row_meta_size = 0, + }, [GGML_TYPE_IQ2_KS] = { .type_name = "iq2_ks", .blck_size = QK_K, @@ -4173,6 +4186,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_IQ4_KS: wtype = GGML_TYPE_IQ4_KS; break; case GGML_FTYPE_MOSTLY_IQ4_KSS: wtype = GGML_TYPE_IQ4_KSS; break; case GGML_FTYPE_MOSTLY_IQ2_K: wtype = GGML_TYPE_IQ2_K; break; + case GGML_FTYPE_MOSTLY_IQ2_K_R4: wtype = GGML_TYPE_IQ2_K_R4; break; 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; @@ -10711,6 +10725,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: case GGML_TYPE_IQ2_K: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: @@ -11168,6 +11183,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: case GGML_TYPE_IQ2_K: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: @@ -11322,6 +11338,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: case GGML_TYPE_IQ2_K: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: @@ -14522,6 +14539,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: case GGML_TYPE_IQ2_K: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: @@ -14916,6 +14934,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: case GGML_TYPE_IQ2_K: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: @@ -15204,6 +15223,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: case GGML_TYPE_IQ2_K: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: @@ -15821,6 +15841,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: case GGML_TYPE_IQ2_K: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: @@ -22666,6 +22687,7 @@ size_t ggml_quantize_chunk( case GGML_TYPE_IQ4_KS: result = quantize_iq4_ks (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_KSS: result = quantize_iq4_kss(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ2_K: result = quantize_iq2_k (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ2_K_R4:result = quantize_iq2_k_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; 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; diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp index bcf96c0a..d08491c3 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -182,6 +182,7 @@ struct MulMat { case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_NL_R4: case GGML_TYPE_IQ4_XS_R4: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ3_K_R4: case GGML_TYPE_IQ4_K_R4: case GGML_TYPE_IQ2_BN_R4: return 4; @@ -3958,6 +3959,108 @@ static void mul_mat_bf16_r16_bf16(int n, const void * vx, size_t bx, const DataI #endif template <int nrc_y> +static void mul_mat_iq2_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 ms = _mm256_set1_epi8(4); + auto m03 = _mm256_set1_epi8(0x03); + auto shift_shuffle = _mm256_set_epi64x(0x0707070706060606, 0x0505050504040404, 0x0303030302020202, 0x0101010100000000); + static const uint8_t kvalues_iq2nl[32] = {1, 19, 33, 49, 6, 24, 38, 54, 1, 19, 33, 49, 6, 24, 38, 54, 1, 19, 33, 49, 6, 24, 38, 54, 1, 19, 33, 49, 6, 24, 38, 54}; + auto values = _mm256_loadu_si256((const __m256i*)kvalues_iq2nl); + 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); +#ifndef HAVE_FANCY_SIMD + auto s_shuffle = _mm256_set_epi64x(0x0f0e0f0e0d0c0d0c, 0x0b0a0b0a09080908, 0x0706070605040504, 0x0302030201000100); +#endif + int nbl = n / QK_K; + __m256 acc[nrc_y] = {}; + __m256i qx[4]; + uint64_t stored_scales[8]; + for (int ix = 0; ix < nrc_x; ix += 4) { + const block_iq2_k_r4 * iq2 = (const block_iq2_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 *)iq2[ibl].d)); + auto d4 = _mm256_set_m128(dl, dl); + auto extra = _mm256_set1_epi64x(*(const uint64_t *)iq2[ibl].extra); + auto slbits = _mm256_loadu_si256((const __m256i *)iq2[ibl].scales); + auto i8scales1 = _mm256_add_epi8(_mm256_and_si256(slbits, m4), _mm256_set1_epi8(-8)); + auto i8scales2 = _mm256_add_epi8(_mm256_and_si256(_mm256_srli_epi16(slbits, 4), m4), _mm256_set1_epi8(-8)); + _mm256_storeu_si256((__m256i *)stored_scales+0, i8scales1); + _mm256_storeu_si256((__m256i *)stored_scales+1, i8scales2); + __m256i isum[nrc_y] = {}; + { + 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_mullo_epi16(_mm256_set1_epi16(-32), MM256_SET_M128I(_mm256_extracti128_si256(t3, 0), _mm256_extracti128_si256(t1, 0))); // blocks 0, 1, 8, 9 + auto s2 = _mm256_mullo_epi16(_mm256_set1_epi16(-32), MM256_SET_M128I(_mm256_extracti128_si256(t3, 1), _mm256_extracti128_si256(t1, 1))); // blocks 2, 3, 10, 11 + auto s3 = _mm256_mullo_epi16(_mm256_set1_epi16(-32), MM256_SET_M128I(_mm256_extracti128_si256(t4, 0), _mm256_extracti128_si256(t2, 0))); // blocks 4, 5, 12, 13 + auto s4 = _mm256_mullo_epi16(_mm256_set1_epi16(-32), 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); +#ifdef HAVE_FANCY_SIMD + isum[iy] = _mm256_dpwssd_epi32(isum[iy], s1, _mm256_shuffle_epi32(bsums, 0x00)); + isum[iy] = _mm256_dpwssd_epi32(isum[iy], s2, _mm256_shuffle_epi32(bsums, 0x55)); + isum[iy] = _mm256_dpwssd_epi32(isum[iy], s3, _mm256_shuffle_epi32(bsums, 0xaa)); + isum[iy] = _mm256_dpwssd_epi32(isum[iy], s4, _mm256_shuffle_epi32(bsums, 0xff)); +#else + isum[iy] = _mm256_add_epi32(isum[iy], _mm256_madd_epi16(s1, _mm256_shuffle_epi32(bsums, 0x00))); + isum[iy] = _mm256_add_epi32(isum[iy], _mm256_madd_epi16(s2, _mm256_shuffle_epi32(bsums, 0x55))); + isum[iy] = _mm256_add_epi32(isum[iy], _mm256_madd_epi16(s3, _mm256_shuffle_epi32(bsums, 0xaa))); + isum[iy] = _mm256_add_epi32(isum[iy], _mm256_madd_epi16(s4, _mm256_shuffle_epi32(bsums, 0xff))); +#endif + } + } + for (int ib = 0; ib < QK_K/32; ++ib) { +#ifdef HAVE_FANCY_SIMD + auto scales = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i *)(stored_scales + ib))); +#else + auto scales = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm_set1_epi64x(stored_scales[ib])), s_shuffle); +#endif + auto lb = _mm256_loadu_si256((const __m256i *)iq2[ibl].qs+ib); + 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_and_si256(lb, m03); + qx[1] = _mm256_and_si256(_mm256_srli_epi16(lb, 2), m03); + qx[2] = _mm256_and_si256(_mm256_srli_epi16(lb, 4), m03); + qx[3] = _mm256_and_si256(_mm256_srli_epi16(lb, 6), m03); + qx[0] = _mm256_shuffle_epi8(values, _mm256_add_epi8(qx[0], shift)); + qx[1] = _mm256_shuffle_epi8(values, _mm256_add_epi8(qx[1], shift)); + qx[2] = _mm256_shuffle_epi8(values, _mm256_add_epi8(qx[2], shift)); + qx[3] = _mm256_shuffle_epi8(values, _mm256_add_epi8(qx[3], shift)); + 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)); + isum[iy] = _mm256_add_epi32(isum[iy], _mm256_mullo_epi32(scales, sumi)); +#else + auto sumi1 = _mm256_add_epi16(_mm256_maddubs_epi16(qx[0], _mm256_shuffle_epi32(y, 0x00)), + _mm256_maddubs_epi16(qx[1], _mm256_shuffle_epi32(y, 0x55))); + auto sumi2 = _mm256_add_epi16(_mm256_maddubs_epi16(qx[2], _mm256_shuffle_epi32(y, 0xaa)), + _mm256_maddubs_epi16(qx[3], _mm256_shuffle_epi32(y, 0xff))); + isum[iy] = _mm256_add_epi32(isum[iy], _mm256_add_epi32(_mm256_madd_epi16(scales, sumi1), _mm256_madd_epi16(scales, sumi2))); +#endif + } + } + for (int iy = 0; iy < nrc_y; ++iy) { + acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(isum[iy]), 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_iq3_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); @@ -6281,6 +6384,18 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) { mm.funcs[7] = mul_mat_iq4_k_r4_q8_k<8>; expected_typeB = GGML_TYPE_Q8_K; break; + case GGML_TYPE_IQ2_K_R4: + assert (ne00 % QK_K == 0); + mm.funcs[0] = mul_mat_iq2_k_r4_q8_k<1>; + mm.funcs[1] = mul_mat_iq2_k_r4_q8_k<2>; + mm.funcs[2] = mul_mat_iq2_k_r4_q8_k<3>; + mm.funcs[3] = mul_mat_iq2_k_r4_q8_k<4>; + mm.funcs[4] = mul_mat_iq2_k_r4_q8_k<5>; + mm.funcs[5] = mul_mat_iq2_k_r4_q8_k<6>; + mm.funcs[6] = mul_mat_iq2_k_r4_q8_k<7>; + mm.funcs[7] = mul_mat_iq2_k_r4_q8_k<8>; + expected_typeB = GGML_TYPE_Q8_K; + break; case GGML_TYPE_IQ3_K_R4: assert (ne00 % QK_K == 0); mm.funcs[0] = mul_mat_iq3_k_r4_q8_k<1>; @@ -8969,11 +9084,19 @@ void mul_mat_iq4_xs_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& i } } -template <int nrc_y> +template <int nrc_y, bool is_iq2k> inline void iq3_4_add_shift(int ibl, const Q8<nrc_y, block_q8_K>& q8, const int8x16x4_t& i8scales, uint8x16_t extra, - uint8x16_t ms, int32x4_t * isum) { - 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)); + int32x4_t * isum) { + auto ms = is_iq2k ? vdupq_n_s8(5) : vdupq_n_s8(4); + int8x16_t s8_1, s8_2; + if constexpr (is_iq2k) { + auto m1 = vdupq_n_u8(1); + s8_1 = vmulq_s8(i8scales.val[0], vandq_s8(ms, vceqq_u8(vandq_u8(extra, m1), m1))); extra = vshrq_n_u8(extra, 2); + s8_2 = vmulq_s8(i8scales.val[1], vandq_s8(ms, vceqq_u8(vandq_u8(extra, m1), m1))); extra = vshrq_n_u8(extra, 2); + } else { + s8_1 = vmulq_s8(i8scales.val[0], vandq_u8(ms, vshlq_n_u8(extra, 2))); + 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)); @@ -8990,8 +9113,14 @@ inline void iq3_4_add_shift(int ibl, const Q8<nrc_y, block_q8_K>& q8, const int8 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))); + if constexpr (is_iq2k) { + auto m1 = vdupq_n_u8(1); + s8_1 = vmulq_s8(i8scales.val[2], vandq_s8(ms, vceqq_u8(vandq_u8(extra, m1), m1))); extra = vshrq_n_u8(extra, 2); + s8_2 = vmulq_s8(i8scales.val[3], vandq_s8(ms, vceqq_u8(vandq_u8(extra, m1), m1))); extra = vshrq_n_u8(extra, 2); + } else { + 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)); @@ -9011,6 +9140,111 @@ inline void iq3_4_add_shift(int ibl, const Q8<nrc_y, block_q8_K>& q8, const int8 } template <int nrc_y> +void mul_mat_iq2_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 m03 = vdupq_n_u8(0x03); + auto ms = vdupq_n_u8(4); + uint8x16x2_t shift_shuffle = { + vreinterpretq_u8_u64(uint64x2_t{0x0101010100000000, 0x0303030302020202}), + vreinterpretq_u8_u64(uint64x2_t{0x0505050504040404, 0x0707070706060606}) + }; + auto values8 = vld1_s8(iq2nl_values); + auto values = vcombine_s8(values8, values8); + 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_iq2_k_r4 * iq2 = (const block_iq2_k_r4 *)((const char *)vx + ix*bx); + for (int ibl = 0; ibl < nbl; ++ibl) { + auto d4 = vcvt_f32_f16(vld1_f16((const float16_t *)iq2[ibl].d)); + auto extra8 = vld1_u8(iq2[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(iq2[ibl].scales); + i8scales.val[0] = vaddq_s8(vandq_u8(sl.val[0], m4), vdupq_n_s8(-8)); + i8scales.val[1] = vaddq_s8(vandq_u8(sl.val[1], m4), vdupq_n_s8(-8)); + i8scales.val[2] = vaddq_s8(vshrq_n_u8(sl.val[0], 4), vdupq_n_s8(-8)); + i8scales.val[3] = vaddq_s8(vshrq_n_u8(sl.val[1], 4), vdupq_n_s8(-8)); + int32x4_t isum[nrc_y] = {}; + if constexpr (nrc_y == 1) { + iq3_4_add_shift<nrc_y, true>(ibl, q8, i8scales, extra, isum); + } + 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 scales = vmovl_s16(vget_low_s16 (i16scales.val[ib])); + auto bits = vld1q_u8_x2(iq2[ibl].qs + 128*is + 32*ib); + qx[0] = vandq_u8( bits.val[0], m03); + qx[1] = vandq_u8(vshrq_n_u8(bits.val[0], 2), m03); + qx[2] = vandq_u8(vshrq_n_u8(bits.val[0], 4), m03); + qx[3] = vandq_u8(vshrq_n_u8(bits.val[0], 6), m03); + uint8x16_t shifts; + if constexpr (nrc_y == 1) { + qx[0] = vqtbl1q_s8(values, qx[0]); // 0...3 from the 4 rows + qx[1] = vqtbl1q_s8(values, qx[1]); // 4...7 + qx[2] = vqtbl1q_s8(values, qx[2]); // 8..11 + qx[3] = vqtbl1q_s8(values, qx[3]); // 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] = vqtbl1q_s8(values, vaddq_u8(shift, qx[0])); // 0...3 from the 4 rows + qx[1] = vqtbl1q_s8(values, vaddq_u8(shift, qx[1])); // 4...7 + qx[2] = vqtbl1q_s8(values, vaddq_u8(shift, qx[2])); // 8..11 + qx[3] = vqtbl1q_s8(values, vaddq_u8(shift, qx[3])); // 12..15 + } + 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); + } + qx[0] = vandq_u8( bits.val[1], m03); + qx[1] = vandq_u8(vshrq_n_u8(bits.val[1], 2), m03); + qx[2] = vandq_u8(vshrq_n_u8(bits.val[1], 4), m03); + qx[3] = vandq_u8(vshrq_n_u8(bits.val[1], 6), m03); + if constexpr (nrc_y == 1) { + qx[0] = vqtbl1q_s8(values, qx[0]); // 0...3 from the 4 rows + qx[1] = vqtbl1q_s8(values, qx[1]); // 4...7 + qx[2] = vqtbl1q_s8(values, qx[2]); // 8..11 + qx[3] = vqtbl1q_s8(values, qx[3]); // 12..15 + } else { + auto shift = vqtbl1q_u8(shifts, shift_shuffle.val[1]); + qx[0] = vqtbl1q_s8(values, vaddq_u8(shift, qx[0])); // 0...3 from the 4 rows + qx[1] = vqtbl1q_s8(values, vaddq_u8(shift, qx[1])); // 4...7 + qx[2] = vqtbl1q_s8(values, vaddq_u8(shift, qx[2])); // 8..11 + qx[3] = vqtbl1q_s8(values, vaddq_u8(shift, qx[3])); // 12..15 + } + 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); + } + } +} + +template <int nrc_y> void mul_mat_iq3_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); @@ -9054,7 +9288,7 @@ void mul_mat_iq3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& in i8scales.val[3] = vmulq_s8(i8scales.val[3], vorrq_u8(vceqq_u8(vandq_u8(sh, smask.val[1]), smask.val[1]), vdupq_n_u8(1))); int32x4_t isum[nrc_y] = {}; if constexpr (nrc_y == 1) { - iq3_4_add_shift(ibl, q8, i8scales, extra, ms, isum); + iq3_4_add_shift<nrc_y, false>(ibl, q8, i8scales, extra, isum); } for (int is = 0; is < 2; ++is) { i16scales.val[0] = vmovl_s8(vget_low_s8 (i8scales.val[2*is+0])); @@ -9161,7 +9395,7 @@ void mul_mat_iq4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& in 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) { - iq3_4_add_shift(ibl, q8, i8scales, extra, ms, isum); + iq3_4_add_shift<nrc_y, false>(ibl, q8, i8scales, extra, isum); } for (int is = 0; is < 2; ++is) { i16scales.val[0] = vmovl_s8(vget_low_s8 (i8scales.val[2*is+0])); @@ -10049,6 +10283,10 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) { SET_MUL_MAT_FUNCTIONS(m, mul_mat_q8_k_r8_q8_k); expected_Btype = GGML_TYPE_Q8_KR8; break; + case GGML_TYPE_IQ2_K_R4: + SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq2_k_r4_q8_k); + expected_Btype = GGML_TYPE_Q8_K; + break; case GGML_TYPE_IQ3_K_R4: SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq3_k_r4_q8_k); expected_Btype = GGML_TYPE_Q8_K; diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp index 373a15bb..3077fe21 100644 --- a/ggml/src/iqk/iqk_quantize.cpp +++ b/ggml/src/iqk/iqk_quantize.cpp @@ -4931,3 +4931,126 @@ void vec_dot_iq3_k_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t GGML_UNUSED(by); } +// +// ========================================= iq2_k_r4 +// + +void quantize_row_iq2_k_r4_ref(const float * x, block_iq2_k_r4 * y, int64_t k) { + quantize_iq2_k_r4(x, (void *)y, 4, k/4, nullptr); +} + +void quantize_row_iq2_k_r4(const float * x, void * y, int64_t k) { + quantize_iq2_k_r4(x, y, 4, k/4, nullptr); +} + +namespace { +inline void convert_iq2_k(const block_iq2_k& x, uint8_t * L) { + const uint8_t * qs = x.qs; + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + int shift_l = 2*(ib32%4); + for (int j = 0; j < 16; ++j) { + L[j+ 0] = ((qs[j+ 0] >> shift_l) & 3); + L[j+16] = ((qs[j+16] >> shift_l) & 3); + } + L += 32; + if (shift_l == 6) qs += 32; + } +} +} + +static void repack_iq2_k(int nrows, int n_per_row, const block_iq2_k * x, block_iq2_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_iq2_k * x4[4]; + uint8_t L[QK_K]; + 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, 0, QK_K/8); + for (int k = 0; k < 4; ++k) { + y[ibl].d[k] = x4[k][ibl].d; + auto extra = x4[k][ibl].extra; + convert_iq2_k(x4[k][ibl], L); + 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[ib] & 0xf; + uint8_t sl2 = x4[k][ibl].scales[ib] >> 4; + int i = 8*ib + k; + y[ibl].scales[i%32] |= (sl1 << 4*(i/32)); + i += 4; + y[ibl].scales[i%32] |= (sl2 << 4*(i/32)); + 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); + } + } + } + } + x += 4*nblock; + y += nblock; + } +} + +size_t quantize_iq2_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_IQ2_K, n_per_row); + std::vector<char> qtmp(4*row_size); + for (int row = 0; row < nrows; row += 4) { + quantize_iq2_k(src, (void *)qtmp.data(), 4, n_per_row, imatrix); + repack_iq2_k(4, n_per_row, (const block_iq2_k *)qtmp.data(), (block_iq2_k_r4 *)qcur); + qcur += 4*row_size; + src += 4*n_per_row; + } + return nrows*row_size; +} + +void dequantize_row_iq2_k_r4(const block_iq2_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; + for (int ib = 0; ib < QK_K/32; ++ib) { + int is = 8*ib + k; + float dl1 = d * (((x[ibl].scales[is%32] >> 4*(is/32)) & 0xf) - 8); + is += 4; + float dl2 = d * (((x[ibl].scales[is%32] >> 4*(is/32)) & 0xf) - 8); + auto values1 = iq2nl_values + (x[ibl].extra[k+0] & (1 << ib) ? 4 : 0); + auto values2 = iq2nl_values + (x[ibl].extra[k+4] & (1 << ib) ? 4 : 0); + for (int i = 0; i < 4; ++i) { + y4[k][QK_K*ibl+32*ib+i+ 0] = dl1 * values1[(ql[4*k+i+ 0] >> 0) & 3]; + y4[k][QK_K*ibl+32*ib+i+ 4] = dl1 * values1[(ql[4*k+i+ 0] >> 2) & 3]; + y4[k][QK_K*ibl+32*ib+i+ 8] = dl1 * values1[(ql[4*k+i+ 0] >> 4) & 3]; + y4[k][QK_K*ibl+32*ib+i+12] = dl1 * values1[(ql[4*k+i+ 0] >> 6) & 3]; + y4[k][QK_K*ibl+32*ib+i+16] = dl2 * values2[(ql[4*k+i+16] >> 0) & 3]; + y4[k][QK_K*ibl+32*ib+i+20] = dl2 * values2[(ql[4*k+i+16] >> 2) & 3]; + y4[k][QK_K*ibl+32*ib+i+24] = dl2 * values2[(ql[4*k+i+16] >> 4) & 3]; + y4[k][QK_K*ibl+32*ib+i+28] = dl2 * values2[(ql[4*k+i+16] >> 6) & 3]; + } + ql += 32; + } + } + } +} + +void vec_dot_iq2_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_IQ2_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 1ca66bd8..8640b59b 100644 --- a/ggml/src/iqk/iqk_quantize.h +++ b/ggml/src/iqk/iqk_quantize.h @@ -151,6 +151,12 @@ size_t quantize_iq3_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT d void dequantize_row_iq3_k_r4(const block_iq3_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); void vec_dot_iq3_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_iq2_k_r4_ref(const float * GGML_RESTRICT x, block_iq2_k_r4 * GGML_RESTRICT y, int64_t k); +void quantize_row_iq2_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +size_t quantize_iq2_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_iq2_k_r4(const block_iq2_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void vec_dot_iq2_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_q8_k_r8_ref(const float * GGML_RESTRICT x, block_q8_k_r8 * GGML_RESTRICT y, int64_t k); void quantize_row_q8_k_r8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); size_t quantize_q8_k_r8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); diff --git a/include/llama.h b/include/llama.h index 026cf08e..1627a752 100644 --- a/include/llama.h +++ b/include/llama.h @@ -193,6 +193,7 @@ extern "C" { LLAMA_FTYPE_MOSTLY_Q6_0_R4 = 335, // except 1d tensors LLAMA_FTYPE_MOSTLY_BF16_R16 = 232, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ2_BN_R4 = 337, // except 1d tensors + LLAMA_FTYPE_MOSTLY_IQ2_K_R4 = 338, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ3_K_R4 = 339, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ4_K_R4 = 340, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q8_K_R8 = 399, // except 1d tensors diff --git a/src/llama.cpp b/src/llama.cpp index 16579e99..62ab4d08 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3866,6 +3866,7 @@ struct llama_model_loader { case GGML_TYPE_IQ4_KS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_KS; break; case GGML_TYPE_IQ4_KSS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_KSS; break; case GGML_TYPE_IQ2_K: ftype = LLAMA_FTYPE_MOSTLY_IQ2_K; break; + case GGML_TYPE_IQ2_K_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ2_K_R4;break; case GGML_TYPE_IQ3_K: ftype = LLAMA_FTYPE_MOSTLY_IQ3_K; break; case GGML_TYPE_IQ3_K_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ3_K_R4;break; case GGML_TYPE_IQ4_K: ftype = LLAMA_FTYPE_MOSTLY_IQ4_K; break; @@ -4585,6 +4586,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ4_KS: return "IQ4_KS - 4.25 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_KSS: return "IQ4_KSS - 4.0 bpw"; case LLAMA_FTYPE_MOSTLY_IQ2_K: return "IQ2_K - 2.375 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ2_K_R4: return "IQ2_K_R4 - 2.375 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_K: return "IQ3_K - 3.4325 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_K_R4: return "IQ3_K_R4 - 3.4325 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_KL: return "IQ3_KL - 4 bpw"; @@ -15765,7 +15767,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K || - ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4) { + ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4) { new_type = !qs.has_output ? GGML_TYPE_IQ4_K : GGML_TYPE_Q5_K; } else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || @@ -15822,6 +15824,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (new_type == GGML_TYPE_Q8_K_R8) { new_type = GGML_TYPE_Q8_0; } + else if (new_type == GGML_TYPE_IQ2_K_R4) { + new_type = GGML_TYPE_IQ2_K; + } else if (new_type == GGML_TYPE_IQ3_K_R4) { new_type = GGML_TYPE_IQ3_K; } @@ -15881,6 +15886,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_K) { new_type = qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ4_K : GGML_TYPE_IQ3_K; } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4) { + new_type = qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ4_K_R4 : GGML_TYPE_IQ3_K_R4; + } else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && qs.model.hparams.n_gqa() >= 4) { new_type = GGML_TYPE_Q4_K; } @@ -16047,7 +16055,8 @@ 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_IQ4_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4) { + ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4|| ftype == LLAMA_FTYPE_MOSTLY_IQ4_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 || + ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4) { new_type = GGML_TYPE_Q5_K; } } else { @@ -16057,6 +16066,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L ) new_type = GGML_TYPE_Q5_K; else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_M ) new_type = GGML_TYPE_IQ4_K; else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_K ) new_type = GGML_TYPE_IQ3_K; + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4) new_type = GGML_TYPE_IQ3_K_R4; else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KL ) new_type = GGML_TYPE_IQ4_KS; } } else { @@ -16118,7 +16128,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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_IQ4_K_R4|| new_type == GGML_TYPE_Q8_K_R8 || new_type == GGML_TYPE_IQ3_K_R4) { + new_type == GGML_TYPE_IQ4_K_R4|| new_type == GGML_TYPE_Q8_K_R8 || new_type == GGML_TYPE_IQ3_K_R4|| + new_type == GGML_TYPE_IQ2_K_R4) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; if (nx % QK_K != 0) { @@ -16149,6 +16160,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n case GGML_TYPE_Q3_K: case GGML_TYPE_Q3_K_R4: case GGML_TYPE_IQ2_K: + case GGML_TYPE_IQ2_K_R4: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ3_K_R4: case GGML_TYPE_IQ4_KSS: @@ -16285,6 +16297,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ4_KS: default_type = GGML_TYPE_IQ4_KS; break; case LLAMA_FTYPE_MOSTLY_IQ4_KSS: default_type = GGML_TYPE_IQ4_KSS; break; case LLAMA_FTYPE_MOSTLY_IQ2_K: default_type = GGML_TYPE_IQ2_K; break; + case LLAMA_FTYPE_MOSTLY_IQ2_K_R4:default_type = GGML_TYPE_IQ2_K_R4;break; case LLAMA_FTYPE_MOSTLY_IQ3_K: default_type = GGML_TYPE_IQ3_K; break; case LLAMA_FTYPE_MOSTLY_IQ3_K_R4:default_type = GGML_TYPE_IQ3_K_R4;break; case LLAMA_FTYPE_MOSTLY_IQ3_KL: default_type = GGML_TYPE_IQ3_K; break; @@ -16695,6 +16708,10 @@ 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_IQ2_K_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_K; + else chunk_size_multiplier = 4; + } else if (new_type == GGML_TYPE_IQ3_K_R4) { if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ3_K; else chunk_size_multiplier = 4; |