diff options
| -rw-r--r-- | examples/quantize/quantize.cpp | 3 | ||||
| -rw-r--r-- | ggml/include/ggml.h | 2 | ||||
| -rw-r--r-- | ggml/src/ggml-common.h | 7 | ||||
| -rw-r--r-- | ggml/src/ggml-quants.c | 1 | ||||
| -rw-r--r-- | ggml/src/ggml.c | 23 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_mul_mat.cpp | 208 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.cpp | 111 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.h | 6 | ||||
| -rw-r--r-- | include/llama.h | 1 | ||||
| -rw-r--r-- | src/llama.cpp | 19 |
10 files changed, 363 insertions, 18 deletions
diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 1ad5108e..dbae9792 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -24,6 +24,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = { { "IQ2_XXS", LLAMA_FTYPE_MOSTLY_IQ2_XXS, " 2.06 bpw quantization", }, { "IQ2_XXS_R4",LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4,"IQ2_XXS repacked", }, { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization", }, + { "IQ2_XS_R4",LLAMA_FTYPE_MOSTLY_IQ2_XS_R4,"IQ2_XS repacked", }, { "IQ2_S", LLAMA_FTYPE_MOSTLY_IQ2_S, " 2.5 bpw quantization", }, { "IQ2_M", LLAMA_FTYPE_MOSTLY_IQ2_M, " 2.7 bpw quantization", }, { "IQ1_S", LLAMA_FTYPE_MOSTLY_IQ1_S, " 1.56 bpw quantization", }, @@ -505,7 +506,7 @@ int main(int argc, char ** argv) { if (!params.ignore_imatrix_rules && imatrix_data.empty() && (params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4 || - params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || + params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS_R4 || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_M)) { fprintf(stderr, "\n==========================================================================================================\n"); diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 42fbcea2..60f787ad 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -419,6 +419,7 @@ extern "C" { GGML_TYPE_Q5_K_R4 = 213, GGML_TYPE_Q6_K_R4 = 214, GGML_TYPE_IQ2_XXS_R4= 216, + GGML_TYPE_IQ2_XS_R4 = 217, GGML_TYPE_IQ3_XXS_R4= 218, GGML_TYPE_IQ4_NL_R4 = 220, GGML_TYPE_IQ4_XS_R4 = 223, @@ -499,6 +500,7 @@ extern "C" { GGML_FTYPE_MOSTLY_Q5_K_R4 = 213, // except 1d tensors GGML_FTYPE_MOSTLY_Q6_K_R4 = 214, // except 1d tensors GGML_FTYPE_MOSTLY_IQ2_XXS_R4= 215, // except 1d tensors + GGML_FTYPE_MOSTLY_IQ2_XS_R4 = 216, // except 1d tensors GGML_FTYPE_MOSTLY_IQ3_XXS_R4= 217, // except 1d tensors GGML_FTYPE_MOSTLY_IQ4_NL_R4 = 219, // except 1d tensors GGML_FTYPE_MOSTLY_IQ4_XS_R4 = 222, // except 1d tensors diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index ad21dd50..2534b461 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -412,6 +412,13 @@ typedef struct { } block_iq2_xs; static_assert(sizeof(block_iq2_xs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding"); +typedef struct { + ggml_half d[4]; + uint16_t qs[QK_K/2]; + uint8_t scales[QK_K/8]; +} block_iq2_xs_r4; +static_assert(sizeof(block_iq2_xs_r4) == 4*sizeof(block_iq2_xs), "wrong iq2_xs_r4 block size/padding"); + // 2.5625 bpw quants typedef struct { ggml_half d; diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index 23c60182..1f56ec06 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -15199,6 +15199,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte case GGML_TYPE_IQ4_NL_R4: break; case GGML_TYPE_IQ4_XS_R4: break; case GGML_TYPE_IQ2_XXS_R4: break; + case GGML_TYPE_IQ2_XS_R4: break; case GGML_TYPE_IQ3_XXS_R4: break; case GGML_TYPE_Q4_0_R4: break; case GGML_TYPE_Q5_0_R4: break; diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 27794cd3..0c3be11c 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1031,6 +1031,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .nrows = 1, .row_meta_size = 0, }, + [GGML_TYPE_IQ2_XS_R4] = { + .type_name = "iq2_xs_r4", + .blck_size = QK_K, + .type_size = sizeof(block_iq2_xs), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_iq2_xs_r4, + .from_float = quantize_row_iq2_xs_r4, + .from_float_ref = (ggml_from_float_t)quantize_row_iq2_xs_r4_ref, + .vec_dot = vec_dot_iq2_xs_r4_q8_k, + .vec_dot_type = GGML_TYPE_Q8_K, + .nrows = 1, + .row_meta_size = 0, + }, [GGML_TYPE_IQ3_XXS] = { .type_name = "iq3_xxs", .blck_size = QK_K, @@ -4226,6 +4239,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_IQ2_XXS: wtype = GGML_TYPE_IQ2_XXS; break; case GGML_FTYPE_MOSTLY_IQ2_XXS_R4: wtype = GGML_TYPE_IQ2_XXS_R4;break; case GGML_FTYPE_MOSTLY_IQ2_XS: wtype = GGML_TYPE_IQ2_XS; break; + case GGML_FTYPE_MOSTLY_IQ2_XS_R4: wtype = GGML_TYPE_IQ2_XS_R4;break; case GGML_FTYPE_MOSTLY_IQ3_XXS: wtype = GGML_TYPE_IQ3_XXS; break; case GGML_FTYPE_MOSTLY_IQ3_XXS_R4: wtype = GGML_TYPE_IQ3_XXS_R4;break; case GGML_FTYPE_MOSTLY_IQ1_S: wtype = GGML_TYPE_IQ1_S; break; @@ -10769,6 +10783,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XXS_R4: case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_XS_R4: case GGML_TYPE_IQ3_XXS: case GGML_TYPE_IQ3_XXS_R4: case GGML_TYPE_IQ1_S: @@ -11231,6 +11246,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XXS_R4: case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_XS_R4: case GGML_TYPE_IQ3_XXS: case GGML_TYPE_IQ3_XXS_R4: case GGML_TYPE_IQ1_S: @@ -11390,6 +11406,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XXS_R4: case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_XS_R4: case GGML_TYPE_IQ3_XXS: case GGML_TYPE_IQ3_XXS_R4: case GGML_TYPE_IQ1_S: @@ -14595,6 +14612,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XXS_R4: case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_XS_R4: case GGML_TYPE_IQ3_XXS: case GGML_TYPE_IQ3_XXS_R4: case GGML_TYPE_IQ1_S: @@ -14994,6 +15012,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XXS_R4: case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_XS_R4: case GGML_TYPE_IQ3_XXS: case GGML_TYPE_IQ3_XXS_R4: case GGML_TYPE_IQ1_S: @@ -15287,6 +15306,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XXS_R4: case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_XS_R4: case GGML_TYPE_IQ3_XXS: case GGML_TYPE_IQ3_XXS_R4: case GGML_TYPE_IQ1_S: @@ -15909,6 +15929,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XXS_R4: case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_XS_R4: case GGML_TYPE_IQ3_XXS: case GGML_TYPE_IQ3_XXS_R4: case GGML_TYPE_IQ1_S: @@ -22680,6 +22701,7 @@ void ggml_quantize_init(enum ggml_type type) { switch (type) { case GGML_TYPE_IQ2_XXS_R4: iq2xs_init_impl(GGML_TYPE_IQ2_XXS); break; + case GGML_TYPE_IQ2_XS_R4: iq2xs_init_impl(GGML_TYPE_IQ2_XS); break; case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XS: case GGML_TYPE_IQ2_S: @@ -22759,6 +22781,7 @@ size_t ggml_quantize_chunk( case GGML_TYPE_IQ2_XXS: result = quantize_iq2_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ2_XXS_R4:result = quantize_iq2_xxs_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ2_XS: result = quantize_iq2_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ2_XS_R4:result = quantize_iq2_xs_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ3_XXS: result = quantize_iq3_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ3_XXS_R4:result = quantize_iq3_xxs_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ3_S: result = quantize_iq3_s (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 0733d4ea..a21700a9 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -3301,6 +3301,131 @@ static void mul_mat_iq2_xxs_r4_q8_k(int n, const void * vx, size_t bx, const Dat } template <int nrc_y> +static void mul_mat_iq2_xs_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); + int nbl = n / QK_K; +#ifndef HAVE_FANCY_SIMD + auto smask = _mm256_set1_epi64x(0x8040201008040201); + auto sign_shuffle = _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202, 0x0101010101010101, 0x0000000000000000); + auto m4 = _mm256_set1_epi8(4); +#endif + __m256 acc[nrc_y] = {}; +#ifdef HAVE_FANCY_SIMD + __m256i shuffles[2] = { + _mm256_set_epi64x(0x0706070607060706, 0x0302030203020302, 0x0504050405040504, 0x0100010001000100), + _mm256_set_epi64x(0x0f0e0f0e0f0e0f0e, 0x0b0a0b0a0b0a0b0a, 0x0d0c0d0c0d0c0d0c, 0x0908090809080908) + }; + __m256i isum[2*nrc_y] = {}; +#else + __m256i shuffles[4] = { + MM256_SET_M128I(_mm_set1_epi16(0x0302), _mm_set1_epi16(0x0100)), + MM256_SET_M128I(_mm_set1_epi16(0x0706), _mm_set1_epi16(0x0504)), + MM256_SET_M128I(_mm_set1_epi16(0x0b0a), _mm_set1_epi16(0x0908)), + MM256_SET_M128I(_mm_set1_epi16(0x0f0e), _mm_set1_epi16(0x0d0c)), + }; + __m256i isum[nrc_y == 1 ? 4 : nrc_y] = {}; +#endif + auto s_shuffle = _mm_set_epi64x(0x0f0d0b0907050301, 0x0e0c0a0806040200); + __m256i qx[4]; + union { __m256i vec; uint16_t val[16]; } helper; + for (int ix = 0; ix < nrc_x; ix += 4) { + auto iq2 = (const block_iq2_xs_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 s32 = (const uint32_t *)iq2[ibl].scales; + for (int ib = 0; ib < QK_K/32; ++ib) { + auto val = _mm256_loadu_si256((const __m256i *)iq2[ibl].qs + ib); + helper.vec = _mm256_and_si256(val, _mm256_set1_epi16(511)); + qx[0] = _mm256_set_epi64x(iq2xs_grid[helper.val[ 3]], iq2xs_grid[helper.val[ 2]], iq2xs_grid[helper.val[ 1]], iq2xs_grid[helper.val[ 0]]); + qx[1] = _mm256_set_epi64x(iq2xs_grid[helper.val[ 7]], iq2xs_grid[helper.val[ 6]], iq2xs_grid[helper.val[ 5]], iq2xs_grid[helper.val[ 4]]); + qx[2] = _mm256_set_epi64x(iq2xs_grid[helper.val[11]], iq2xs_grid[helper.val[10]], iq2xs_grid[helper.val[ 9]], iq2xs_grid[helper.val[ 8]]); + qx[3] = _mm256_set_epi64x(iq2xs_grid[helper.val[15]], iq2xs_grid[helper.val[14]], iq2xs_grid[helper.val[13]], iq2xs_grid[helper.val[12]]); + auto signs16 = _mm256_srli_epi16(val, 9); + signs16 = _mm256_xor_si256(signs16, _mm256_slli_epi16(signs16, 1)); + auto signs128 = _mm_or_si128(_mm256_castsi256_si128(signs16), _mm_slli_epi16(_mm256_extracti128_si256(signs16, 1), 8)); + signs128 = _mm_shuffle_epi8(signs128, s_shuffle); + auto scales = _mm_set1_epi32(s32[ib]); + scales = _mm_and_si128(_mm_unpacklo_epi8(scales, _mm_srli_epi16(scales, 4)), _mm_set1_epi8(0xf)); + scales = _mm_or_si128(_mm_slli_epi16(scales, 1), _mm_set1_epi8(1)); + auto scales16 = _mm256_cvtepi8_epi16(scales); // 0...7, 0...7 +#ifdef HAVE_FANCY_SIMD + __m256i scs[2] = { _mm256_shuffle_epi8(scales16, shuffles[0]), _mm256_shuffle_epi8(scales16, shuffles[1]) }; + auto mask = (const __mmask32 *)&signs128; + for (int iy = 0; iy < nrc_y; ++iy) { + auto y = _mm256_loadu_si256((const __m256i *)q8.y[iy][ibl].qs + ib); + auto sumi1 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[0], _mm256_mask_sub_epi8(y, mask[0], _mm256_setzero_si256(), y)); // blocks: 0,0,0,0, 1,1,1,1, row 0 + auto sumi2 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[1], _mm256_mask_sub_epi8(y, mask[1], _mm256_setzero_si256(), y)); // blocks: 2,2,2,2, 3,3,3,3, row 1 + auto sumi3 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[2], _mm256_mask_sub_epi8(y, mask[2], _mm256_setzero_si256(), y)); // blocks: 4,4,4,4, 5,5,5,5, row 2 + auto sumi4 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[3], _mm256_mask_sub_epi8(y, mask[3], _mm256_setzero_si256(), y)); // blocks: 6,6,6,6, 7,7,7,7, row 3 + auto s12 = _mm256_packs_epi32(sumi1, sumi2); // 0,0,0,0, 2,2,2,2, 1,1,1,1, 3,3,3,3 + auto s34 = _mm256_packs_epi32(sumi3, sumi4); // 4,4,4,4, 6,6,6,6, 5,5,5,5, 7,7,7,7 + isum[2*iy+0] = _mm256_add_epi32(isum[2*iy+0], _mm256_madd_epi16(scs[0], s12)); + isum[2*iy+1] = _mm256_add_epi32(isum[2*iy+1], _mm256_madd_epi16(scs[1], s34)); + } +#else + auto signs = MM256_SET_M128I(signs128, signs128); + auto shuffle = sign_shuffle; + auto s1 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1)); + shuffle = _mm256_add_epi8(shuffle, m4); + auto s2 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1)); + shuffle = _mm256_add_epi8(shuffle, m4); + auto s3 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1)); + shuffle = _mm256_add_epi8(shuffle, m4); + auto s4 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1)); + __m256i scs[4] = { + _mm256_shuffle_epi8(scales16, shuffles[0]), _mm256_shuffle_epi8(scales16, shuffles[1]), + _mm256_shuffle_epi8(scales16, shuffles[2]), _mm256_shuffle_epi8(scales16, shuffles[3]), + }; + for (int iy = 0; iy < nrc_y; ++iy) { + auto y = _mm256_loadu_si256((const __m256i *)q8.y[iy][ibl].qs + ib); + if constexpr (nrc_y == 1) { + isum[0] = _mm256_add_epi32(isum[0], _mm256_madd_epi16(scs[0], _mm256_maddubs_epi16(qx[0], _mm256_sign_epi8(y, s1)))); + isum[1] = _mm256_add_epi32(isum[1], _mm256_madd_epi16(scs[1], _mm256_maddubs_epi16(qx[1], _mm256_sign_epi8(y, s2)))); + isum[2] = _mm256_add_epi32(isum[2], _mm256_madd_epi16(scs[2], _mm256_maddubs_epi16(qx[2], _mm256_sign_epi8(y, s3)))); + isum[3] = _mm256_add_epi32(isum[3], _mm256_madd_epi16(scs[3], _mm256_maddubs_epi16(qx[3], _mm256_sign_epi8(y, s4)))); + } else { + auto sumi1 = _mm256_madd_epi16(scs[0], _mm256_maddubs_epi16(qx[0], _mm256_sign_epi8(y, s1))); // blocks 4x0, 4x1, row 0 + auto sumi2 = _mm256_madd_epi16(scs[1], _mm256_maddubs_epi16(qx[1], _mm256_sign_epi8(y, s2))); // blocks 4x2, 4x3, row 1 + auto sumi3 = _mm256_madd_epi16(scs[2], _mm256_maddubs_epi16(qx[2], _mm256_sign_epi8(y, s3))); // blocks 4x4, 4x5, row 2 + auto sumi4 = _mm256_madd_epi16(scs[3], _mm256_maddubs_epi16(qx[3], _mm256_sign_epi8(y, s4))); // blocks 4x6, 4x7, row 3 + auto s12 = _mm256_add_epi32(_mm256_unpacklo_epi32(sumi1, sumi2), _mm256_unpackhi_epi32(sumi1, sumi2)); // 0,1, 0,1, 0,1, 0,1 + auto s34 = _mm256_add_epi32(_mm256_unpacklo_epi32(sumi3, sumi4), _mm256_unpackhi_epi32(sumi3, sumi4)); // 2,3, 2,3, 2,3, 2,3 + auto sumi = _mm256_add_epi32(_mm256_unpacklo_epi64(s12, s34), _mm256_unpackhi_epi64(s12, s34)); // 0,1,2,3, 0,1,2,3 + isum[iy] = _mm256_add_epi32(isum[iy], sumi); + } + } +#endif + } + for (int iy = 0; iy < nrc_y; ++iy) { +#ifdef HAVE_FANCY_SIMD + auto sumi = _mm256_hadd_epi32(isum[2*iy+0], isum[2*iy+1]); + acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(sumi), acc[iy]); + isum[2*iy+0] = isum[2*iy+1] = _mm256_setzero_si256(); +#else + if constexpr (nrc_y == 1) { + auto s12 = _mm256_add_epi32(_mm256_unpacklo_epi32(isum[0], isum[1]), _mm256_unpackhi_epi32(isum[0], isum[1])); + auto s34 = _mm256_add_epi32(_mm256_unpacklo_epi32(isum[2], isum[3]), _mm256_unpackhi_epi32(isum[2], isum[3])); + auto sumi = _mm256_add_epi32(_mm256_unpacklo_epi64(s12, s34), _mm256_unpackhi_epi64(s12, s34)); + acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(sumi), acc[iy]); + isum[0] = isum[1] = isum[2] = isum[3] = _mm256_setzero_si256(); + } else { + acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(isum[iy]), acc[iy]); + isum[iy] = _mm256_setzero_si256(); + } +#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)); + info.store(ix, iy, _mm_mul_ps(_mm_set1_ps(0.125f), sum)); + acc[iy] = _mm256_setzero_ps(); + } + } +} + +template <int nrc_y> static void mul_mat_iq3_xxs_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); @@ -6801,6 +6926,18 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) { mm.funcs[7] = mul_mat_iq2_xxs_r4_q8_k<8>; expected_typeB = GGML_TYPE_Q8_K; break; + case GGML_TYPE_IQ2_XS_R4: + assert (ne00 % QK_K == 0); + mm.funcs[0] = mul_mat_iq2_xs_r4_q8_k<1>; + mm.funcs[1] = mul_mat_iq2_xs_r4_q8_k<2>; + mm.funcs[2] = mul_mat_iq2_xs_r4_q8_k<3>; + mm.funcs[3] = mul_mat_iq2_xs_r4_q8_k<4>; + mm.funcs[4] = mul_mat_iq2_xs_r4_q8_k<5>; + mm.funcs[5] = mul_mat_iq2_xs_r4_q8_k<6>; + mm.funcs[6] = mul_mat_iq2_xs_r4_q8_k<7>; + mm.funcs[7] = mul_mat_iq2_xs_r4_q8_k<8>; + expected_typeB = GGML_TYPE_Q8_K; + break; case GGML_TYPE_IQ3_XXS_R4: assert (ne00 % QK_K == 0); mm.funcs[0] = mul_mat_iq3_xxs_r4_q8_k<1>; @@ -9735,6 +9872,73 @@ static void mul_mat_iq2_xxs_r4_q8_k(int n, const void * vx, size_t bx, const Dat } template <int nrc_y> +static void mul_mat_iq2_xs_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); + int nbl = n / QK_K; + static const uint8_t k_shuff[16] = {1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31}; + auto shuff = vld1q_u8(k_shuff); + float32x4_t acc[nrc_y] = {}; + int32x4_t isum[2*nrc_y] = {}; + int8x16_t qx[8]; + uint16x8x4_t scales16; + SignHelper sh; + for (int ix = 0; ix < nrc_x; ix += 4) { + auto iq2 = (const block_iq2_xs_r4 *)((const char *)vx + (ix+0)*bx); + for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256 + auto d4 = vcvt_f32_f16(vld1_f16((const float16_t *)iq2[ibl].d)); + auto qs = iq2[ibl].qs; + for (int is = 0; is < 2; ++is) { + auto scale_bits = vld1q_u8(iq2[ibl].scales + 16*is); + auto scales1 = vandq_u8(scale_bits, vdupq_n_u8(0xf)); + auto scales2 = vshrq_n_u8(scale_bits, 4); + scales1 = vorrq_u8(vshlq_n_u8(scales1, 1), vdupq_n_u8(1)); + scales2 = vorrq_u8(vshlq_n_u8(scales2, 1), vdupq_n_u8(1)); + auto s1 = vzip1q_u8(scales1, scales2); + auto s2 = vzip2q_u8(scales1, scales2); + scales16.val[0] = vmovl_u8(vget_low_u8 (s1)); + scales16.val[1] = vmovl_u8(vget_high_u8(s1)); + scales16.val[2] = vmovl_u8(vget_low_u8 (s2)); + scales16.val[3] = vmovl_u8(vget_high_u8(s2)); + for (int ib = 0; ib < QK_K/64; ++ib) { + auto v = vld1q_u8_x2((const uint8_t *)qs); + auto signs128 = vandq_u8(vqtbl2q_u8(v, shuff), vdupq_n_u8(254)); + signs128 = veorq_u8(signs128, vshrq_n_u8(signs128, 1)); + sh.init(); + for (int i = 0; i < 8; ++i) { + qx[i] = vreinterpretq_s8_u64(uint64x2_t{iq2xs_grid[qs[2*i+0] & 511], iq2xs_grid[qs[2*i+1] & 511]}); + sh.apply_signs_1((uint8x16_t *)qx+i, signs128); + } + auto s32_1 = vmovl_u16(vget_low_u16 (scales16.val[ib])); + auto s32_2 = vmovl_u16(vget_high_u16(scales16.val[ib])); + for (int iy = 0; iy < nrc_y; ++iy) { + auto y = vld1q_s8_x2(q8.y[iy][ibl].qs + 128*is + 32*ib); + auto sumi1 = vpaddq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[0], y.val[0]), ggml_vdotq_s32(vdupq_n_s32(0), qx[1], y.val[1])); + auto sumi2 = vpaddq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[2], y.val[0]), ggml_vdotq_s32(vdupq_n_s32(0), qx[3], y.val[1])); + auto sumi3 = vpaddq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[4], y.val[0]), ggml_vdotq_s32(vdupq_n_s32(0), qx[5], y.val[1])); + auto sumi4 = vpaddq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[6], y.val[0]), ggml_vdotq_s32(vdupq_n_s32(0), qx[7], y.val[1])); + auto sumi12 = vpaddq_s32(sumi1, sumi2); // blocks 0,1,2,3 in rows 0,1 + auto sumi34 = vpaddq_s32(sumi3, sumi4); // blocks 4,5,6,7 in rows 2,3 + isum[2*iy+0] = vmlaq_s32(isum[2*iy+0], s32_1, sumi12); + isum[2*iy+1] = vmlaq_s32(isum[2*iy+1], s32_2, sumi34); + } + qs += 16; + } + } + for (int iy = 0; iy < nrc_y; ++iy) { + auto sumi = vpaddq_s32(isum[2*iy+0], isum[2*iy+1]); + acc[iy] = vfmaq_f32(acc[iy], vmulq_f32(d4, vdupq_n_f32(q8.scale(iy, ibl))), vcvtq_f32_s32(sumi)); + isum[2*iy] = isum[2*iy+1] = vdupq_n_s32(0); + } + } + for (int iy = 0; iy < nrc_y; ++iy) { + info.store(ix, iy, vmulq_f32(vdupq_n_f32(0.125f), acc[iy])); + acc[iy] = vdupq_n_f32(0.f); + } + } +} + +template <int nrc_y> static void mul_mat_iq3_xxs_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); @@ -11085,6 +11289,10 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) { SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq2_xxs_r4_q8_k); expected_Btype = GGML_TYPE_Q8_K; break; + case GGML_TYPE_IQ2_XS_R4: + SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq2_xs_r4_q8_k); + expected_Btype = GGML_TYPE_Q8_K; + break; case GGML_TYPE_IQ3_XXS_R4: SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq3_xxs_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 4c49836e..e369a2f0 100644 --- a/ggml/src/iqk/iqk_quantize.cpp +++ b/ggml/src/iqk/iqk_quantize.cpp @@ -5303,18 +5303,6 @@ struct Repack { }; } -// -// ========================================= iq2_xxs_r4 -// - -void quantize_row_iq2_xxs_r4_ref(const float * x, block_iq2_xxs_r4 * y, int64_t k) { - quantize_iq2_xxs_r4(x, (void *)y, 4, k/4, nullptr); -} - -void quantize_row_iq2_xxs_r4(const float * x, void * y, int64_t k) { - quantize_iq2_xxs_r4(x, y, 4, k/4, nullptr); -} - namespace { inline uint8_t scrambled_sign(uint8_t s) { static const uint8_t k_table[128] = { @@ -5331,6 +5319,18 @@ inline uint8_t scrambled_sign(uint8_t s) { } } +// +// ========================================= iq2_xxs_r4 +// + +void quantize_row_iq2_xxs_r4_ref(const float * x, block_iq2_xxs_r4 * y, int64_t k) { + quantize_iq2_xxs_r4(x, (void *)y, 4, k/4, nullptr); +} + +void quantize_row_iq2_xxs_r4(const float * x, void * y, int64_t k) { + quantize_iq2_xxs_r4(x, y, 4, k/4, nullptr); +} + static void repack_iq2_xxs(int nrows, int n_per_row, const block_iq2_xxs * x, block_iq2_xxs_r4 * y) { GGML_ASSERT(nrows%4 == 0); GGML_ASSERT(n_per_row%QK_K == 0); @@ -5420,6 +5420,93 @@ void vec_dot_iq2_xxs_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_ } // +// ========================================= iq2_xs_r4 +// + +void quantize_row_iq2_xs_r4_ref(const float * x, block_iq2_xs_r4 * y, int64_t k) { + quantize_iq2_xs_r4(x, (void *)y, 4, k/4, nullptr); +} + +void quantize_row_iq2_xs_r4(const float * x, void * y, int64_t k) { + quantize_iq2_xs_r4(x, y, 4, k/4, nullptr); +} + +static void repack_iq2_xs(int nrows, int n_per_row, const block_iq2_xs * x, block_iq2_xs_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_xs * 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) { + for (int k = 0; k < 4; ++k) { + y[ibl].d[k] = x4[k][ibl].d; + for (int ib = 0; ib < QK_K/32; ++ib) { + for (int i = 0; i < 4; ++i) { + uint16_t v = x4[k][ibl].qs[4*ib+i]; + uint8_t s = v >> 9; + y[ibl].qs[16*ib+4*k+i] = (v & 511) | (scrambled_sign(s) << 9); + } + y[ibl].scales[4*ib+k] = x4[k][ibl].scales[ib]; + } + } + } + x += 4*nblock; + y += nblock; + } +} + +size_t quantize_iq2_xs_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_XS, n_per_row); + std::vector<char> qtmp(4*row_size); + for (int row = 0; row < nrows; row += 4) { + quantize_iq2_xs(src, (void *)qtmp.data(), 4, n_per_row, imatrix); + repack_iq2_xs(4, n_per_row, (const block_iq2_xs *)qtmp.data(), (block_iq2_xs_r4 *)qcur); + qcur += 4*row_size; + src += 4*n_per_row; + } + return nrows*row_size; +} + +void dequantize_row_iq2_xs_r4(const block_iq2_xs_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 = 0.125f*GGML_FP16_TO_FP32(x[ibl].d[k]); + for (int ib = 0; ib < QK_K/32; ++ib) { + float dl1 = d * (2*(x[ibl].scales[4*ib+k] & 0xf) + 1); + float dl2 = d * (2*(x[ibl].scales[4*ib+k] >> 4) + 1); + for (int i = 0; i < 4; ++i) { + auto val = (const int8_t *)(iq2xs_grid + (x[ibl].qs[16*ib+4*k+i] & 511)); + auto signs = x[ibl].qs[16*ib+4*k+i] >> 9; + signs ^= (signs << 1); + float dl = i < 2 ? dl1 : dl2; + for (int j = 0; j < 8; ++j) y4[k][QK_K*ibl+32*ib+8*i+j] = dl * val[j] * (signs & (1 << j) ? -1 : 1); + } + } + } + } +} + +void vec_dot_iq2_xs_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_XS_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); +} + +// // ========================================= iq3_xxs_r4 // diff --git a/ggml/src/iqk/iqk_quantize.h b/ggml/src/iqk/iqk_quantize.h index 18fc0773..7b183956 100644 --- a/ggml/src/iqk/iqk_quantize.h +++ b/ggml/src/iqk/iqk_quantize.h @@ -175,6 +175,12 @@ size_t quantize_iq2_xxs_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT void dequantize_row_iq2_xxs_r4(const block_iq2_xxs_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); void vec_dot_iq2_xxs_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_xs_r4_ref(const float * GGML_RESTRICT x, block_iq2_xs_r4 * GGML_RESTRICT y, int64_t k); +void quantize_row_iq2_xs_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +size_t quantize_iq2_xs_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_xs_r4(const block_iq2_xs_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void vec_dot_iq2_xs_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_iq3_xxs_r4_ref(const float * GGML_RESTRICT x, block_iq3_xxs_r4 * GGML_RESTRICT y, int64_t k); void quantize_row_iq3_xxs_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); size_t quantize_iq3_xxs_r4(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 27d10c14..8c8fbe6a 100644 --- a/include/llama.h +++ b/include/llama.h @@ -189,6 +189,7 @@ extern "C" { LLAMA_FTYPE_MOSTLY_Q5_K_R4 = 216, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q6_K_R4 = 218, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4 = 219, // except 1d tensors + LLAMA_FTYPE_MOSTLY_IQ2_XS_R4 = 220, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 = 223, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 = 225, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ4_XS_R4 = 230, // except 1d tensors diff --git a/src/llama.cpp b/src/llama.cpp index df700c12..eac0d866 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3852,6 +3852,7 @@ struct llama_model_loader { case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break; case GGML_TYPE_IQ2_XXS_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4; break; case GGML_TYPE_IQ2_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS; break; + case GGML_TYPE_IQ2_XS_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS_R4; break; case GGML_TYPE_IQ2_KS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_KS; break; case GGML_TYPE_IQ2_S: ftype = LLAMA_FTYPE_MOSTLY_IQ2_S; break; case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break; @@ -4581,6 +4582,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ2_XXS: return "IQ2_XXS - 2.0625 bpw"; case LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4:return "IQ2_XXS_R4 - 2.0625 bpw"; case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ2_XS_R4:return "IQ2_XS_R4 - 2.3125 bpw"; case LLAMA_FTYPE_MOSTLY_IQ2_KS: return "IQ2_KS - 2.1875 bpw"; case LLAMA_FTYPE_MOSTLY_IQ2_S: return "IQ2_S - 2.5 bpw"; case LLAMA_FTYPE_MOSTLY_IQ2_M: return "IQ2_M - 2.7 bpw"; @@ -15797,10 +15799,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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_K_R4 || - ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) { + ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) { new_type = !qs.has_output ? GGML_TYPE_IQ4_K : GGML_TYPE_Q5_K; } - else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4) { + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS_R4) { new_type = !qs.has_output ? GGML_TYPE_IQ4_K_R4 : GGML_TYPE_Q5_K_R4; } else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || @@ -15818,7 +15820,7 @@ 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_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || - ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4) { + ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS_R4) { new_type = GGML_TYPE_Q2_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) { @@ -15894,7 +15896,7 @@ 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_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_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4) { + ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS_R4) { if (name.find("attn_v.weight") != std::string::npos) { if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_IQ4_K; else if (qs.model.hparams.n_gqa() >= 2 || qs.model.hparams.n_expert >= 2) new_type = GGML_TYPE_IQ3_K; @@ -16188,7 +16190,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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_IQ2_K_R4|| new_type == GGML_TYPE_IQ5_K_R4|| new_type == GGML_TYPE_IQ4_KS_R4 || - new_type == GGML_TYPE_IQ3_XXS_R4 || new_type == GGML_TYPE_IQ2_XXS_R4) { + new_type == GGML_TYPE_IQ3_XXS_R4 || new_type == GGML_TYPE_IQ2_XXS_R4 || new_type == GGML_TYPE_IQ2_XS_R4) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; if (nx % QK_K != 0) { @@ -16209,6 +16211,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XXS_R4: case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ2_XS_R4: case GGML_TYPE_IQ2_KS: case GGML_TYPE_IQ2_S: case GGML_TYPE_IQ3_XXS: @@ -16341,6 +16344,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ2_XXS: default_type = GGML_TYPE_IQ2_XXS; break; case LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4:default_type = GGML_TYPE_IQ2_XXS_R4; break; case LLAMA_FTYPE_MOSTLY_IQ2_XS: default_type = GGML_TYPE_IQ2_XS; break; + case LLAMA_FTYPE_MOSTLY_IQ2_XS_R4:default_type = GGML_TYPE_IQ2_XS_R4; break; case LLAMA_FTYPE_MOSTLY_IQ2_KS: default_type = GGML_TYPE_IQ2_KS; break; case LLAMA_FTYPE_MOSTLY_IQ2_S: default_type = GGML_TYPE_IQ2_XS; break; case LLAMA_FTYPE_MOSTLY_IQ2_M: default_type = GGML_TYPE_IQ2_S; break; @@ -16695,6 +16699,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s (new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_XXS_R4 || new_type == GGML_TYPE_IQ2_XS || + new_type == GGML_TYPE_IQ2_XS_R4 || new_type == GGML_TYPE_IQ2_S || new_type == GGML_TYPE_IQ1_S || (new_type == GGML_TYPE_IQ1_M && strcmp(tensor->name, "token_embd.weight") && strcmp(tensor->name, "output.weight")) || @@ -16800,6 +16805,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_XXS; else chunk_size_multiplier = 4; } + else if (new_type == GGML_TYPE_IQ2_XS_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_XS; + else chunk_size_multiplier = 4; + } else if (new_type == GGML_TYPE_IQ3_XXS_R4) { if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ3_XXS; else chunk_size_multiplier = 4; |