diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index ec45e96d8f..94e09840ed 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -386,45 +386,44 @@ extern "C" { // NOTE: always add types at the end of the enum to keep backward compatibility enum ggml_type { - GGML_TYPE_F32 = 0, - GGML_TYPE_F16 = 1, - GGML_TYPE_Q4_0 = 2, - GGML_TYPE_Q4_1 = 3, - // GGML_TYPE_Q4_2 = 4, support has been removed - // GGML_TYPE_Q4_3 = 5, support has been removed - GGML_TYPE_Q5_0 = 6, - GGML_TYPE_Q5_1 = 7, - GGML_TYPE_Q8_0 = 8, - GGML_TYPE_Q8_1 = 9, - GGML_TYPE_Q2_K = 10, - GGML_TYPE_Q3_K = 11, - GGML_TYPE_Q4_K = 12, - GGML_TYPE_Q5_K = 13, - GGML_TYPE_Q6_K = 14, - GGML_TYPE_Q8_K = 15, - GGML_TYPE_IQ2_XXS = 16, - GGML_TYPE_IQ2_XS = 17, - GGML_TYPE_IQ3_XXS = 18, - GGML_TYPE_IQ1_S = 19, - GGML_TYPE_IQ4_NL = 20, - GGML_TYPE_IQ3_S = 21, - GGML_TYPE_IQ2_S = 22, - GGML_TYPE_IQ4_XS = 23, - GGML_TYPE_I8 = 24, - GGML_TYPE_I16 = 25, - GGML_TYPE_I32 = 26, - GGML_TYPE_I64 = 27, - GGML_TYPE_F64 = 28, - GGML_TYPE_IQ1_M = 29, - GGML_TYPE_BF16 = 30, - GGML_TYPE_Q4_0_4_4 = 31, - GGML_TYPE_Q4_0_4_8 = 32, - GGML_TYPE_Q4_0_8_8 = 33, - GGML_TYPE_MXFP4 = 39, // so we are compatible with mainline + GGML_TYPE_F32 = 0, + GGML_TYPE_F16 = 1, + GGML_TYPE_Q4_0 = 2, + GGML_TYPE_Q4_1 = 3, + // GGML_TYPE_Q4_2 = 4, support has been removed + // GGML_TYPE_Q4_3 = 5, support has been removed + GGML_TYPE_Q5_0 = 6, + GGML_TYPE_Q5_1 = 7, + GGML_TYPE_Q8_0 = 8, + GGML_TYPE_Q8_1 = 9, + GGML_TYPE_Q2_K = 10, + GGML_TYPE_Q3_K = 11, + GGML_TYPE_Q4_K = 12, + GGML_TYPE_Q5_K = 13, + GGML_TYPE_Q6_K = 14, + GGML_TYPE_Q8_K = 15, + GGML_TYPE_IQ2_XXS = 16, + GGML_TYPE_IQ2_XS = 17, + GGML_TYPE_IQ3_XXS = 18, + GGML_TYPE_IQ1_S = 19, + GGML_TYPE_IQ4_NL = 20, + GGML_TYPE_IQ3_S = 21, + GGML_TYPE_IQ2_S = 22, + GGML_TYPE_IQ4_XS = 23, + GGML_TYPE_I8 = 24, + GGML_TYPE_I16 = 25, + GGML_TYPE_I32 = 26, + GGML_TYPE_I64 = 27, + GGML_TYPE_F64 = 28, + GGML_TYPE_IQ1_M = 29, + GGML_TYPE_BF16 = 30, + GGML_TYPE_Q4_0_4_4 = 31, + GGML_TYPE_Q4_0_4_8 = 32, + GGML_TYPE_Q4_0_8_8 = 33, + GGML_TYPE_I2_S = 36, // So we are able to consume MS BitNet I2_S quants + GGML_TYPE_MXFP4 = 39, // so we are compatible with mainline + GGML_TYPE_Q1_0_G128 = 41, // Bonsai 1-bit quants // - // So we are able to consume MS BitNet I2_S quants - // - GGML_TYPE_I2_S = 36, // GGML_TYPE_Q8_0_X4 = 97, GGML_TYPE_Q8_1_X4 = 98, @@ -530,6 +529,7 @@ extern "C" { GGML_FTYPE_MOSTLY_Q4_0_4_4 = 26, // except 1d tensors GGML_FTYPE_MOSTLY_Q4_0_4_8 = 27, // except 1d tensors GGML_FTYPE_MOSTLY_Q4_0_8_8 = 28, // except 1d tensors + GGML_FTYPE_MOSTLY_Q1_0_128 = 29, // except 1d tensors // GGML_FTYPE_MOSTLY_Q6_0 = 127, // except 1d tensors GGML_FTYPE_MOSTLY_IQ1_BN = 128, // except 1d tensors diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index a59b5a0204..708f46f257 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -541,6 +541,16 @@ typedef struct { } block_iq1_m_r4; static_assert(sizeof(block_iq1_m_r4) == 28, "wrong iq1_m_r4 block size/padding"); +// +// Bonsai +// +#define QK1_0_G128 128 +typedef struct { + ggml_half d; + uint8_t qs[QK1_0_G128 / 8]; +} block_q1_0_g128; +static_assert(sizeof(block_q1_0_g128) == sizeof(ggml_half) + QK1_0_G128 / 8, "wrong q1_0_g128 block size/padding"); + // // Bitnet and TriLM - implemented as 1.625 bpw // diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index 0f3caece74..6f9007e079 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -15435,6 +15435,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte case GGML_TYPE_IQ2_KT: break; case GGML_TYPE_IQ3_KT: break; case GGML_TYPE_IQ4_KT: break; + case GGML_TYPE_Q1_0_G128: break; case GGML_TYPE_IQ3_K: break; case GGML_TYPE_IQ3_KS: break; case GGML_TYPE_IQ2_KL: break; diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index b7a4a14e44..7b67b600ac 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1677,6 +1677,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .nrows = 1, .row_meta_size = 4, }, + [GGML_TYPE_Q1_0_G128] = { + .type_name = "q1_0_g128", + .blck_size = QK1_0_G128, + .type_size = sizeof(block_q1_0_g128), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_q1_0_g128, + .from_float = quantize_row_q1_0_g128, + .from_float_ref = (ggml_from_float_t)quantize_row_q1_0_g128_ref, + .vec_dot = vec_dot_q1_0_g128_q8_0, + .vec_dot_type = GGML_TYPE_Q8_0_X4, + .nrows = 1, + .row_meta_size = 0, + }, [GGML_TYPE_IQ3_K] = { .type_name = "iq3_k", .blck_size = QK_K, @@ -4900,6 +4913,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_IQ2_KT: wtype = GGML_TYPE_IQ2_KT; break; case GGML_FTYPE_MOSTLY_IQ3_KT: wtype = GGML_TYPE_IQ3_KT; break; case GGML_FTYPE_MOSTLY_IQ4_KT: wtype = GGML_TYPE_IQ4_KT; break; + case GGML_FTYPE_MOSTLY_Q1_0_128: wtype = GGML_TYPE_Q1_0_G128;break; case GGML_FTYPE_MOSTLY_IQ3_K: wtype = GGML_TYPE_IQ3_K; break; case GGML_FTYPE_MOSTLY_IQ3_KS: wtype = GGML_TYPE_IQ3_KS; break; case GGML_FTYPE_MOSTLY_IQ2_KL: wtype = GGML_TYPE_IQ2_KL; break; @@ -12817,6 +12831,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_IQ2_KT: case GGML_TYPE_IQ3_KT: case GGML_TYPE_IQ4_KT: + case GGML_TYPE_Q1_0_G128: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ3_KS: case GGML_TYPE_IQ2_KL: @@ -13370,6 +13385,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_IQ2_KT: case GGML_TYPE_IQ3_KT: case GGML_TYPE_IQ4_KT: + case GGML_TYPE_Q1_0_G128: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ3_KS: case GGML_TYPE_IQ2_KL: @@ -13549,6 +13565,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_IQ2_KT: case GGML_TYPE_IQ3_KT: case GGML_TYPE_IQ4_KT: + case GGML_TYPE_Q1_0_G128: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ3_KS: case GGML_TYPE_IQ2_KL: @@ -17874,6 +17891,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_IQ2_KT: case GGML_TYPE_IQ3_KT: case GGML_TYPE_IQ4_KT: + case GGML_TYPE_Q1_0_G128: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ3_KS: case GGML_TYPE_IQ2_KL: @@ -18297,6 +18315,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_IQ2_KT: case GGML_TYPE_IQ3_KT: case GGML_TYPE_IQ4_KT: + case GGML_TYPE_Q1_0_G128: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ3_KS: case GGML_TYPE_IQ2_KL: @@ -18626,6 +18645,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_IQ2_KT: case GGML_TYPE_IQ3_KT: case GGML_TYPE_IQ4_KT: + case GGML_TYPE_Q1_0_G128: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ3_KS: case GGML_TYPE_IQ2_KL: @@ -19383,6 +19403,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_IQ2_KT: case GGML_TYPE_IQ3_KT: case GGML_TYPE_IQ4_KT: + case GGML_TYPE_Q1_0_G128: case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ3_KS: case GGML_TYPE_IQ2_KL: @@ -28464,6 +28485,7 @@ size_t ggml_quantize_chunk( case GGML_TYPE_IQ2_KT: result = quantize_iq2_kt (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ3_KT: result = quantize_iq3_kt (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_KT: result = quantize_iq4_kt (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q1_0_G128: result = quantize_q1_0_g128(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_IQ3_KS: result = quantize_iq3_ks (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ2_KL: result = quantize_iq2_kl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; diff --git a/ggml/src/iqk/iqk_gemm_1bit.cpp b/ggml/src/iqk/iqk_gemm_1bit.cpp index 33363c8ed3..35d6ca99fe 100644 --- a/ggml/src/iqk/iqk_gemm_1bit.cpp +++ b/ggml/src/iqk/iqk_gemm_1bit.cpp @@ -1445,6 +1445,63 @@ IQK_NOINLINE void mul_mat_iq2bn_q8_K64(int n, const void * vx, size_t bx, const } } +template +static void mul_mat_q1_0_g128_q8_0(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { + constexpr int n4 = QK1_0_G128 / QK8_0; + Q8 q8(info); + const block_q8_0_x4 * y[nrc_y]; + for (int iy = 0; iy < nrc_y; ++iy) { + y[iy] = (const block_q8_0_x4 *)info.src1_row(iy); + } + __m256i shuffle[4] = { + _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202, 0x0101010101010101, 0x0000000000000000), + _mm256_set_epi64x(0x0707070707070707, 0x0606060606060606, 0x0505050505050505, 0x0404040404040404), + _mm256_set_epi64x(0x0b0b0b0b0b0b0b0b, 0x0a0a0a0a0a0a0a0a, 0x0909090909090909, 0x0808080808080808), + _mm256_set_epi64x(0x0f0f0f0f0f0f0f0f, 0x0e0e0e0e0e0e0e0e, 0x0d0d0d0d0d0d0d0d, 0x0c0c0c0c0c0c0c0c), + }; + auto mask = _mm256_set1_epi64x(0x8040201008040201); + auto mp1 = _mm256_set1_epi8( 1); + auto mm1 = _mm256_set1_epi8(-1); + auto m1 = _mm256_set1_epi16(1); + int nb = n / QK1_0_G128; + __m256i qx[4]; + __m256i sumi[4]; + for (int ix = 0; ix < nrc_x; ++ix) { + auto x = (const block_q1_0_g128 *)((const char *)vx + ix*bx); + __m256 acc[nrc_y] = {}; + for (int ib = 0; ib < nb; ++ib) { + float d = GGML_FP16_TO_FP32(x[ib].d); + auto vd = _mm256_set1_ps(d); + auto bits128 = _mm_loadu_si128((const __m128i *)x[ib].qs); + auto bits = MM256_SET_M128I(bits128, bits128); + for (int k = 0; k < 4; ++k) { + qx[k] = _mm256_shuffle_epi8(bits, shuffle[k]); + qx[k] = _mm256_cmpeq_epi8(_mm256_and_si256(qx[k], mask), mask); + qx[k] = _mm256_or_si256(_mm256_and_si256(qx[k], mp1), _mm256_andnot_si256(qx[k], mm1)); + } + for (int iy = 0; iy < nrc_y; ++iy) { + for (int k = 0; k < n4; ++k) { + auto qy = _mm256_loadu_si256((const __m256i *)y[iy][ib].qs + k); +#ifdef HAVE_VNNI256 + sumi[k] = _mm256_dpbusd_epi32(_mm256_setzero_si256(), mp1, _mm256_sign_epi8(qy, qx[k])); +#else + sumi[k] = _mm256_madd_epi16(m1, _mm256_maddubs_epi16(mp1, _mm256_sign_epi8(qy, qx[k]))); +#endif + } + sumi[0] = _mm256_madd_epi16(m1, _mm256_packs_epi32(sumi[0], sumi[1])); + sumi[2] = _mm256_madd_epi16(m1, _mm256_packs_epi32(sumi[2], sumi[3])); + sumi[0] = _mm256_madd_epi16(m1, _mm256_packs_epi32(sumi[0], sumi[2])); + auto dy = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)y[iy][ib].d)); + auto dxy= _mm256_mul_ps(vd, _mm256_set_m128(dy, dy)); + acc[iy] = _mm256_fmadd_ps(dxy, _mm256_cvtepi32_ps(sumi[0]), acc[iy]); + } + } + for (int iy = 0; iy < nrc_y; ++iy) { + info.store(ix, iy, hsum_float_8(acc[iy])); + } + } +} + template static void mul_mat_iq2_bn_r4_q8_k16_avx2(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { if (nrc_x%4) { @@ -1903,6 +1960,11 @@ bool iqk_set_kernels_1bit(int ne00, int typeA, int typeB, std::array +static void mul_mat_q1_0_g128_q8_0(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { + Q8 q8(info); + for (int ix = 0; ix < nrc_x; ++ix) { + auto x = (const block_q1_0_g128 *)((const char *)vx + ix*bx); + for (int iy = 0; iy < nrc_y; ++iy) { + float s; + vec_dot_q1_0_g128_q8_0(n, &s, 0, x, bx, q8.y[iy], 0, 1); + info.store(ix, iy, s); + } + } +} + template static void mul_mat_iq1_s_r4_q8_1(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { GGML_ASSERT(nrc_x%4 == 0); @@ -2831,6 +2906,11 @@ bool iqk_set_kernels_1bit(int ne00, int typeA, int typeB, std::array; expected_Btype = GGML_TYPE_Q8_K128; break; + case GGML_TYPE_Q1_0_G128: + if (ne00 % QK1_0_G128 != 0) return false; + expected_Btype = GGML_TYPE_Q8_0_X4; + IQK_SET_MUL_MAT_FUNCTIONS(mul_mat_q1_0_g128_q8_0, funcs); + break; default: return false; } diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp index 1ee910db87..607c68d829 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -941,6 +941,7 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) { case GGML_TYPE_IQ1_BN: case GGML_TYPE_IQ2_BN: case GGML_TYPE_IQ2_BN_R4: + case GGML_TYPE_Q1_0_G128: return iqk_set_kernels_1bit(ne00, typeA, typeB, mm.funcs, mm.func16); default: @@ -1032,6 +1033,7 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) { case GGML_TYPE_IQ1_M: case GGML_TYPE_IQ1_S_R4: case GGML_TYPE_IQ1_M_R4: + case GGML_TYPE_Q1_0_G128: return iqk_set_kernels_1bit(ne00, typeA, typeB, m.funcs, m.func16); case GGML_TYPE_IQ1_KT: case GGML_TYPE_IQ2_KT: diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp index 86a082230f..07edaf0392 100644 --- a/ggml/src/iqk/iqk_quantize.cpp +++ b/ggml/src/iqk/iqk_quantize.cpp @@ -9872,6 +9872,91 @@ void vec_dot_iq4_kt_q8_k(int n, float * s, size_t bs, const void * vx, size_t bx } +void quantize_row_q1_0_g128_ref(const float * x, block_q1_0_g128 * y, int64_t k) { + quantize_row_q1_0_g128(x, y, k); +} + +void quantize_row_q1_0_g128(const float * x, void * vy, int64_t k) { + assert(k % QK1_0_G128 == 0); + int nb = k / QK1_0_G128; + auto y = (block_q1_0_g128 *)vy; + for (int ib = 0; ib < nb; ++ib) { + float sum = 0; + for (int j = 0; j < QK1_0_G128; ++j) sum += std::abs(x[j]); + float d = sum / QK1_0_G128; + y[ib].d = GGML_FP32_TO_FP16(d); + std::memset(y[ib].qs, 0, QK1_0_G128/8); + for (int j = 0; j < QK1_0_G128; ++j) { + if (x[j] >= 0.0f) { + y[ib].qs[j / 8] |= (1 << (j % 8)); + } + } + x += QK1_0_G128; + } +} + +size_t quantize_q1_0_g128(const float * src, void * dst, int64_t nrows, int64_t n_per_row, [[maybe_unused]] const float * imatrix) { + GGML_ASSERT(n_per_row % QK1_0_G128 == 0); + int64_t ntot = nrows * n_per_row; + quantize_row_q1_0_g128(src, dst, ntot); + int64_t nblock = ntot / QK1_0_G128; + return nblock * sizeof(block_q1_0_g128); +} + +void dequantize_row_q1_0_g128(const block_q1_0_g128 * x, float * y, int64_t k) { + assert(k % QK1_0_G128 == 0); + constexpr uint8_t k_mask[8] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80}; + int nb = k / QK1_0_G128; + for (int ib = 0; ib < nb; ++ib) { + float d = GGML_FP16_TO_FP32(x[ib].d); + for (int i = 0; i < QK1_0_G128/8; ++i) { + for (int j = 0; j < 8; ++j) { + *y++ = x[ib].qs[i] & k_mask[j] ? d : -d; + } + } + } +} + +void vec_dot_q1_0_g128_q8_0(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) { + assert(n % QK1_0_G128 == 0); + assert(nrc == 1); + GGML_UNUSED(nrc); + GGML_UNUSED(bx); + GGML_UNUSED(by); + GGML_UNUSED(bs); + int nb = n / QK1_0_G128; + + constexpr uint8_t k_mask[8] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80}; + + constexpr int n4 = QK1_0_G128 / QK8_0; + + auto x = (const block_q1_0_g128 *)vx; + auto y = (const block_q8_0_x4 *)vy; + int16_t sumi[QK1_0_G128/8]; + float sumf = 0; + for (int ib = 0; ib < nb; ++ib) { + auto dx = GGML_FP16_TO_FP32(x[ib].d); + auto qx = x[ib].qs; + auto qy = y[ib].qs; + for (int k = 0; k < QK1_0_G128/8; ++k) { + uint8_t bits = qx[k]; + int16_t s = 0; + for (int j = 0; j < 8; ++j) { + s += (bits & k_mask[j] ? qy[j] : -qy[j]); + } + qy += 8; + sumi[k] = s; + } + auto s = sumi; + for (int k = 0; k < n4; ++k) { + float dy = GGML_FP16_TO_FP32(y[ib].d[k]); + sumf += dx*dy*(s[0] + s[1] + s[2] + s[3]); + s += 4; + } + } + *s = sumf; +} + namespace { template inline int check_row_for_blocks_256_fp16(int nblock, const Block * x) { diff --git a/ggml/src/iqk/iqk_quantize.h b/ggml/src/iqk/iqk_quantize.h index 6e4cc5870a..d85c75996f 100644 --- a/ggml/src/iqk/iqk_quantize.h +++ b/ggml/src/iqk/iqk_quantize.h @@ -295,6 +295,12 @@ size_t quantize_q8_KV_r8(const float * GGML_RESTRICT src, void * GGML_RESTRICT d void dequantize_row_q8_KV_r8(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); void vec_dot_q8_KV_r8_q8_KV(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_q1_0_g128_ref(const float * GGML_RESTRICT x, block_q1_0_g128 * GGML_RESTRICT y, int64_t k); +void quantize_row_q1_0_g128(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +size_t quantize_q1_0_g128(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +void dequantize_row_q1_0_g128(const block_q1_0_g128 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void vec_dot_q1_0_g128_q8_0(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 a2f044098a..44817d9f75 100644 --- a/include/llama.h +++ b/include/llama.h @@ -189,6 +189,7 @@ extern "C" { LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // except 1d tensors LLAMA_FTYPE_MOSTLY_MXFP4 = 38, // except 1d tensors, 38 to be compatible with mainline + LLAMA_FTYPE_MOSTLY_Q1_0_G128 = 41, // except 1d tensors, 38 to be compatible with mainline // LLAMA_FTYPE_MOSTLY_Q6_0 = 135, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ1_BN = 136, // except 1d tensors diff --git a/src/llama-model.cpp b/src/llama-model.cpp index af587fd3de..2d5f483e7e 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -1678,6 +1678,7 @@ std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_Q6_0_R4: return "Q6_0_R4 - 6.5 bpw"; case LLAMA_FTYPE_MOSTLY_Q8_0_R8: return "Q8_0_R8 - 8.5 bpw"; case LLAMA_FTYPE_MOSTLY_MXFP4: return "MXFP4 - 4.25 bpw"; + case LLAMA_FTYPE_MOSTLY_Q1_0_G128:return "Q1_0_G128 - 1.125 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_KS: return "IQ4_KS - 4.25 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_KS_R4:return "IQ4_KS_R4 - 4.25 bpw"; diff --git a/src/llama-quantize.cpp b/src/llama-quantize.cpp index 49c3605149..56a94c3962 100644 --- a/src/llama-quantize.cpp +++ b/src/llama-quantize.cpp @@ -988,6 +988,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_Q6_0_R4: default_type = GGML_TYPE_Q6_0_R4; break; case LLAMA_FTYPE_MOSTLY_Q8_0_R8: default_type = GGML_TYPE_Q8_0_R8; break; case LLAMA_FTYPE_MOSTLY_MXFP4: default_type = GGML_TYPE_MXFP4; break; + case LLAMA_FTYPE_MOSTLY_Q1_0_G128: default_type = GGML_TYPE_Q1_0_G128; break; case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break; case LLAMA_FTYPE_MOSTLY_IQ4_KS: default_type = GGML_TYPE_IQ4_KS; break; case LLAMA_FTYPE_MOSTLY_IQ4_KS_R4:default_type = GGML_TYPE_IQ4_KS_R4;break;