From 24d3bf2e70f9821e2056bf6c7c7da98e0af8670f Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Sun, 8 Dec 2024 19:48:15 +0200 Subject: [PATCH] Something is still wrong --- examples/quantize/quantize.cpp | 1 + ggml/include/ggml.h | 2 + ggml/src/ggml-common.h | 8 +++ ggml/src/ggml-quants.c | 1 + ggml/src/ggml.c | 26 ++++++- ggml/src/iqk/iqk_mul_mat.cpp | 97 ++++++++++++++++++++++++++ ggml/src/iqk/iqk_quantize.cpp | 121 +++++++++++++++++++++++++++++++++ ggml/src/iqk/iqk_quantize.h | 6 ++ include/llama.h | 1 + src/llama.cpp | 18 ++++- 10 files changed, 276 insertions(+), 5 deletions(-) diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 24e5a61b..731ea3ff 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -58,6 +58,7 @@ static const std::vector QUANT_OPTIONS = { { "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", }, + { "Q4_K_R4", LLAMA_FTYPE_MOSTLY_Q4_K_R4, "Q4_K_S repacked", }, { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", }, { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", }, { "Q5_K", LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", }, diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index bb50a025..38e01dd5 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -412,6 +412,7 @@ extern "C" { GGML_TYPE_Q4_0_R4 = 202, GGML_TYPE_Q5_0_R4 = 206, GGML_TYPE_Q8_0_R4 = 208, + GGML_TYPE_Q4_K_R4 = 212, GGML_TYPE_IQ4_NL_R4 = 220, GGML_TYPE_IQ4_XS_R4 = 223, GGML_TYPE_Q6_0_R4 = 233, @@ -478,6 +479,7 @@ extern "C" { GGML_FTYPE_MOSTLY_Q4_0_R4 = 202, // except 1d tensors GGML_FTYPE_MOSTLY_Q8_0_R4 = 207, // except 1d tensors GGML_FTYPE_MOSTLY_Q5_0_R4 = 208, // except 1d tensors + GGML_FTYPE_MOSTLY_Q4_K_R4 = 212, // except 1d tensors GGML_FTYPE_MOSTLY_IQ4_NL_R4 = 219, // except 1d tensors GGML_FTYPE_MOSTLY_IQ4_XS_R4 = 222, // except 1d tensors GGML_FTYPE_MOSTLY_Q6_0_R4 = 227, // except 1d tensors diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index 3f701873..784125fc 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -299,6 +299,14 @@ typedef struct { } block_q4_K; static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding"); +typedef struct { + ggml_half d[8]; + uint8_t scales_h[QK_K/16];// scales and mins, quantized with 6 bits + uint8_t scales_l[QK_K/8]; // scales and mins, quantized with 6 bits + uint8_t qs[QK_K*2]; // 4--bit quants +} block_q4_k_r4; +static_assert(sizeof(block_q4_k_r4) == 8*sizeof(ggml_half) + QK_K/16 + QK_K/8 + QK_K*2, "wrong q4_k_r4 block size/padding"); + // 5-bit quantization // 8 blocks of 32 elements each // weight is represented as x = a * q + b diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index 7eece2b3..a4b234c5 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -15202,6 +15202,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte case GGML_TYPE_Q5_0_R4: break; case GGML_TYPE_Q6_0_R4: break; case GGML_TYPE_Q8_0_R4: break; + case GGML_TYPE_Q4_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 974e42b2..fadda3e3 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -888,6 +888,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .nrows = 1, .row_meta_size = 0, }, + [GGML_TYPE_Q4_K_R4] = { + .type_name = "q4_k_r4", + .blck_size = QK_K, + .type_size = sizeof(block_q4_K), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_q4_k_r4, + .from_float = quantize_row_q4_k_r4, + .from_float_ref = (ggml_from_float_t) quantize_row_q4_k_r4_ref, + .vec_dot = vec_dot_q4_k_r4_q8_k, + .vec_dot_type = GGML_TYPE_Q8_K32, + .nrows = 1, + .row_meta_size = 0, + }, [GGML_TYPE_Q5_K] = { .type_name = "q5_K", .blck_size = QK_K, @@ -4020,8 +4033,9 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_Q2_K: wtype = GGML_TYPE_Q2_K; break; case GGML_FTYPE_MOSTLY_Q3_K: wtype = GGML_TYPE_Q3_K; break; case GGML_FTYPE_MOSTLY_Q4_K: wtype = GGML_TYPE_Q4_K; break; - case GGML_FTYPE_MOSTLY_Q5_K: wtype = GGML_TYPE_Q5_K; break; - case GGML_FTYPE_MOSTLY_Q6_K: wtype = GGML_TYPE_Q6_K; break; + case GGML_FTYPE_MOSTLY_Q4_K_R4: wtype = GGML_TYPE_Q4_K_R4; break; + case GGML_FTYPE_MOSTLY_Q5_K: wtype = GGML_TYPE_Q5_K; break; + case GGML_FTYPE_MOSTLY_Q6_K: wtype = GGML_TYPE_Q6_K; break; case GGML_FTYPE_MOSTLY_IQ2_XXS: wtype = GGML_TYPE_IQ2_XXS; break; case GGML_FTYPE_MOSTLY_IQ2_XS: wtype = GGML_TYPE_IQ2_XS; break; case GGML_FTYPE_MOSTLY_IQ3_XXS: wtype = GGML_TYPE_IQ3_XXS; break; @@ -10550,6 +10564,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_Q2_K: case GGML_TYPE_Q3_K: case GGML_TYPE_Q4_K: + case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: @@ -10999,6 +11014,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_Q2_K: case GGML_TYPE_Q3_K: case GGML_TYPE_Q4_K: + case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: @@ -11145,6 +11161,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_Q2_K: case GGML_TYPE_Q3_K: case GGML_TYPE_Q4_K: + case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: @@ -14337,6 +14354,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_Q2_K: case GGML_TYPE_Q3_K: case GGML_TYPE_Q4_K: + case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: @@ -14723,6 +14741,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_Q2_K: case GGML_TYPE_Q3_K: case GGML_TYPE_Q4_K: + case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: @@ -15003,6 +15022,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_Q2_K: case GGML_TYPE_Q3_K: case GGML_TYPE_Q4_K: + case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: @@ -15610,6 +15630,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_Q2_K: case GGML_TYPE_Q3_K: case GGML_TYPE_Q4_K: + case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: case GGML_TYPE_IQ2_XXS: @@ -22445,6 +22466,7 @@ size_t ggml_quantize_chunk( case GGML_TYPE_Q2_K: result = quantize_q2_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q3_K: result = quantize_q3_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q4_K: result = quantize_q4_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q4_K_R4: result = quantize_q4_k_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q5_K: result = quantize_q5_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q6_K: result = quantize_q6_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ2_XXS: result = quantize_iq2_xxs(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 a9adedfc..c6dd1ed7 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -3073,6 +3073,91 @@ static void mul_mat_iq4_xs_r4_q8_k(int n, const void * vx, size_t bx, const Data } #endif +template +static void mul_mat_q4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { + //if constexpr (nrc_y == 1){ + // mul_mat_iq4_xs_r4_q8_k_avx2<1>(n, vx, bx, info, nrc_x); + //} else { + GGML_ASSERT(nrc_x%8 == 0); + Q8 q8(info); + auto mf = _mm512_set1_epi8(0xf); + int nbl = n / QK_K; + using helper_t = union { __m512i vec; uint32_t val[16]; }; + helper_t hd, hm; + __m512 acc[2*nrc_y] = {}; + __m512i qx[4]; + for (int ix = 0; ix < nrc_x; ix += 8) { + const block_q4_k_r4 * iq4l = (const block_q4_k_r4 *)((const char *)vx + (ix+0)*bx); + const block_q4_k_r4 * iq4h = (const block_q4_k_r4 *)((const char *)vx + (ix+4)*bx); + for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256 + auto d1 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)iq4l[ibl].d)); + auto d2 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)iq4h[ibl].d)); + auto dl = _mm256_castps256_ps128(d1); + auto ml = _mm256_extractf128_ps(d1, 1); + auto dh = _mm256_castps256_ps128(d2); + auto mh = _mm256_extractf128_ps(d2, 1); + auto d4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_set_m128(dl, dl)), _mm256_set_m128(dh, dh), 1); + auto m4 = _mm512_insertf32x8(_mm512_castps256_ps512(_mm256_set_m128(ml, ml)), _mm256_set_m128(mh, mh), 1); + auto slbits_l = _mm256_loadu_si256((const __m256i *)iq4l[ibl].scales_l); + auto shbits_l = _mm256_loadu_si256((const __m256i *)iq4h[ibl].scales_l); + auto slb = _mm512_inserti32x8(_mm512_castsi256_si512(slbits_l), shbits_l, 1); + auto sld = _mm512_and_si512(slb, mf); + auto slm = _mm512_and_si512(_mm512_srli_epi16(slb, 4), mf); + auto slbits_h = _mm_loadu_si128((const __m128i *)iq4l[ibl].scales_h); + auto shbits_h = _mm_loadu_si128((const __m128i *)iq4h[ibl].scales_h); + auto slbits_h2 = MM256_SET_M128I(_mm_srli_epi16(slbits_h, 4), slbits_h); + auto shbits_h2 = MM256_SET_M128I(_mm_srli_epi16(shbits_h, 4), shbits_h); + auto shb = _mm512_inserti32x8(_mm512_castsi256_si512(slbits_h2), shbits_h2, 1); + auto shd = _mm512_and_si512(_mm512_slli_epi16(shb, 4), _mm512_set1_epi8(0x30)); + auto shm = _mm512_and_si512(_mm512_slli_epi16(shb, 2), _mm512_set1_epi8(0x30)); + hd.vec = _mm512_or_si512(sld, shd); + hm.vec = _mm512_or_si512(slm, shm); + for (int ib = 0; ib < QK_K/32; ++ib) { + //auto iscales = _mm512_cvtepi8_epi32(_mm_set_epi32(hd.val[ib+8], hd.val[ib+8], hd.val[ib], hd.val[ib])); + auto scales1 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hd.val[ib+0])); + auto scales2 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hd.val[ib+8])); + auto iscales = _mm512_inserti32x8(_mm512_castsi256_si512(scales1), scales2, 1); + auto scales = _mm512_mul_ps(d4, _mm512_cvtepi32_ps(iscales)); + //iscales = _mm512_cvtepi8_epi32(_mm_set_epi32(hm.val[ib+8], hm.val[ib+8], hm.val[ib], hm.val[ib])); + scales1 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hm.val[ib+0])); + scales2 = _mm256_cvtepi8_epi32(_mm_set1_epi32(hm.val[ib+8])); + iscales = _mm512_inserti32x8(_mm512_castsi256_si512(scales1), scales2, 1); + auto scales_m = _mm512_mul_ps(m4, _mm512_cvtepi32_ps(iscales)); + auto bits1 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq4l[ibl].qs+2*ib+0)), + _mm256_loadu_si256((const __m256i *)iq4h[ibl].qs+2*ib+0), 1); + auto bits2 = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)iq4l[ibl].qs+2*ib+1)), + _mm256_loadu_si256((const __m256i *)iq4h[ibl].qs+2*ib+1), 1); + qx[0] = _mm512_and_si512(bits1, mf); + qx[1] = _mm512_and_si512(bits2, mf); + qx[2] = _mm512_and_si512(_mm512_srli_epi16(bits1, 4), mf); + qx[3] = _mm512_and_si512(_mm512_srli_epi16(bits2, 4), mf); + for (int iy = 0; iy < nrc_y; ++iy) { + auto y8 = _mm256_loadu_si256((const __m256i*)q8.y[iy][ibl].qs+ib); + auto y = _mm512_inserti32x8(_mm512_castsi256_si512(y8), y8, 1); + auto sumi = _mm512_setzero_si512(); + sumi = _mm512_dpbusd_epi32(sumi, qx[0], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0x00))); + sumi = _mm512_dpbusd_epi32(sumi, qx[1], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0x55))); + sumi = _mm512_dpbusd_epi32(sumi, qx[2], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xaa))); + sumi = _mm512_dpbusd_epi32(sumi, qx[3], _mm512_shuffle_epi32(y, _MM_PERM_ENUM(0xff))); + float d8 = q8.scale(iy, ibl); + float m8 = ((const float *)q8.y[iy][ibl].bsums)[ib]; + acc[2*iy+0] = _mm512_fmadd_ps(_mm512_mul_ps(scales, _mm512_set1_ps(d8)), _mm512_cvtepi32_ps(sumi), acc[2*iy+0]); + acc[2*iy+1] = _mm512_fmadd_ps(scales_m, _mm512_set1_ps(m8), acc[2*iy+1]); + } + } + } + for (int iy = 0; iy < nrc_y; ++iy) { + auto sum512 = _mm512_fmadd_ps(_mm512_set1_ps(-0.5f), acc[2*iy+1], acc[2*iy+0]); + acc[2*iy+0] = acc[2*iy+1] = _mm512_setzero_ps(); + auto sum1 = _mm_add_ps(_mm512_extractf32x4_ps(sum512, 0), _mm512_extractf32x4_ps(sum512, 1)); + auto sum2 = _mm_add_ps(_mm512_extractf32x4_ps(sum512, 2), _mm512_extractf32x4_ps(sum512, 3)); + info.store(ix+0, iy, sum1); + info.store(ix+4, iy, sum2); + } + } + //} +} + template inline void multiply_add_1(int j, const Bits& bits, const __m256i * scales, const __m256i * q8, __m256i * sumi) { if (j == 0) { @@ -5068,6 +5153,18 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) { mm.funcs[7] = mul_mat_iq4_xs_r4_q8_k<8>; expected_typeB = GGML_TYPE_Q8_K32; break; + case GGML_TYPE_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_Q4_0_R4: assert (ne00 % QK4_NL == 0); mm.funcs[0] = mul_mat_q4_0_r4_q8_1<1>; diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp index 66cf32bc..6084d97e 100644 --- a/ggml/src/iqk/iqk_quantize.cpp +++ b/ggml/src/iqk/iqk_quantize.cpp @@ -3942,3 +3942,124 @@ void vec_dot_iq2_bn_r4_q8_K64(int n, float * s, size_t bs, const void * vx, size GGML_UNUSED(by); } +// +// ========================================= q4_k_r4 +// + +void quantize_row_q4_k_r4_ref(const float * x, block_q4_k_r4 * y, int64_t k) { + quantize_q4_k_r4(x, (void *)y, 4, k/4, nullptr); +} + +void quantize_row_q4_k_r4(const float * x, void * y, int64_t k) { + quantize_q4_k_r4(x, y, 4, k/4, nullptr); +} + +namespace { +inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t& d, uint8_t& m) { + if (j < 4) { + d = q[j] & 63; m = q[j + 4] & 63; + } else { + d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4); + m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); + } +} +} + +static void repack_q4_k(int nrows, int n_per_row, const block_q4_K * x, block_q4_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_q4_K * x4[4]; + uint8_t ld, lm; + for (int row = 0; row < nrows; row += 4) { + for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k; + for (int ibl = 0; ibl < nblock; ++ibl) { + std::memset(y[ibl].scales_l, 0, QK_K/8); + std::memset(y[ibl].scales_h, 0, QK_K/16); + for (int k = 0; k < 4; ++k) { + y[ibl].d[k+0] = x4[k][ibl].d; + y[ibl].d[k+4] = x4[k][ibl].dmin; + for (int ib = 0; ib < QK_K/32; ++ib) { + get_scale_min_k4(ib, x4[k][ibl].scales, ld, lm); + y[ibl].scales_l[4*ib+k] = (ld & 0xf) | ((lm & 0xf) << 4); + uint8_t h = (ld >> 4) | ((lm >> 4) << 2); + y[ibl].scales_h[(4*ib+k)%16] |= (h << 4*((4*ib+k)/16)); + } + } + for (int ib = 0; ib < QK_K/32; ++ib) { + for (int k = 0; k < 4; ++k) for (int i = 0; i < 4; ++i) { + uint8_t l1 = (x4[k][ibl].qs[32*(ib/2)+i+ 0] >> 4*(ib%2)) & 0xf; + uint8_t l2 = (x4[k][ibl].qs[32*(ib/2)+i+ 8] >> 4*(ib%2)) & 0xf; + y[ibl].qs[64*ib+4*k+i+ 0] = l1 | (l2 << 4); + l1 = (x4[k][ibl].qs[32*(ib/2)+i+16] >> 4*(ib%2)) & 0xf; + l2 = (x4[k][ibl].qs[32*(ib/2)+i+24] >> 4*(ib%2)) & 0xf; + y[ibl].qs[64*ib+4*k+i+16] = l1 | (l2 << 4); + l1 = (x4[k][ibl].qs[32*(ib/2)+i+ 4] >> 4*(ib%2)) & 0xf; + l2 = (x4[k][ibl].qs[32*(ib/2)+i+12] >> 4*(ib%2)) & 0xf; + y[ibl].qs[64*ib+4*k+i+32] = l1 | (l2 << 4); + l1 = (x4[k][ibl].qs[32*(ib/2)+i+20] >> 4*(ib%2)) & 0xf; + l2 = (x4[k][ibl].qs[32*(ib/2)+i+28] >> 4*(ib%2)) & 0xf; + y[ibl].qs[64*ib+4*k+i+48] = l1 | (l2 << 4); + } + } + } + x += 4*nblock; + y += nblock; + } +} + +size_t quantize_q4_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_Q4_K, n_per_row); + std::vector qtmp(4*row_size); + for (int row = 0; row < nrows; row += 4) { + quantize_q4_K(src, (void *)qtmp.data(), 4, n_per_row, imatrix); + repack_q4_k(4, n_per_row, (const block_q4_K *)qtmp.data(), (block_q4_k_r4 *)qcur); + qcur += 4*row_size; + src += 4*n_per_row; + } + return nrows*row_size; +} + +void dequantize_row_q4_k_r4(const block_q4_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+0]); + const float m = GGML_FP16_TO_FP32(x[ibl].d[k+4]); + for (int ib = 0; ib < QK_K/32; ++ib) { + int is = 4*ib + k; + float dl = d * ((x[ibl].scales_l[is] & 0xf) | (((x[ibl].scales_h[is%16] >> 4*(is/16)) & 0x03) << 4)); + float ml = m * ((x[ibl].scales_l[is] >> 4) | (((x[ibl].scales_h[is%16] >> 4*(is/16)) & 0x0c) << 2)); + for (int i = 0; i < 4; ++i) { + y4[k][QK_K*ibl+32*ib+i+ 0] = dl * (x[ibl].qs[64*ib+4*k+i+ 0] & 0xf) - ml; + y4[k][QK_K*ibl+32*ib+i+ 8] = dl * (x[ibl].qs[64*ib+4*k+i+ 0] >> 4) - ml; + y4[k][QK_K*ibl+32*ib+i+16] = dl * (x[ibl].qs[64*ib+4*k+i+16] & 0xf) - ml; + y4[k][QK_K*ibl+32*ib+i+24] = dl * (x[ibl].qs[64*ib+4*k+i+16] >> 4) - ml; + y4[k][QK_K*ibl+32*ib+i+ 4] = dl * (x[ibl].qs[64*ib+4*k+i+32] & 0xf) - ml; + y4[k][QK_K*ibl+32*ib+i+12] = dl * (x[ibl].qs[64*ib+4*k+i+32] >> 4) - ml; + y4[k][QK_K*ibl+32*ib+i+20] = dl * (x[ibl].qs[64*ib+4*k+i+48] & 0xf) - ml; + y4[k][QK_K*ibl+32*ib+i+28] = dl * (x[ibl].qs[64*ib+4*k+i+48] >> 4) - ml; + } + } + } + } +} + +void vec_dot_q4_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_Q4_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 c370c8e8..5b4a3e44 100644 --- a/ggml/src/iqk/iqk_quantize.h +++ b/ggml/src/iqk/iqk_quantize.h @@ -109,6 +109,12 @@ void dequantize_row_iq2_bn_r4(const block_iq2_bn * GGML_RESTRICT x, float * GG size_t quantize_iq2_bn_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); void vec_dot_iq2_bn_r4_q8_K64(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); +void quantize_row_q4_k_r4_ref(const float * GGML_RESTRICT x, block_q4_k_r4 * GGML_RESTRICT y, int64_t k); +void quantize_row_q4_k_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +size_t quantize_q4_k_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +void dequantize_row_q4_k_r4(const block_q4_k_r4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void vec_dot_q4_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 9eec6a43..2fa78879 100644 --- a/include/llama.h +++ b/include/llama.h @@ -183,6 +183,7 @@ extern "C" { LLAMA_FTYPE_MOSTLY_Q4_0_R4 = 202, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q8_0_R4 = 207, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q5_0_R4 = 208, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_K_R4 = 214, // 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 diff --git a/src/llama.cpp b/src/llama.cpp index 4b25d4ec..daa270b0 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -4546,6 +4546,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium"; case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "Q3_K - Large"; case LLAMA_FTYPE_MOSTLY_Q4_K_S: return "Q4_K - Small"; + case LLAMA_FTYPE_MOSTLY_Q4_K_R4: return "Q4_K_R4"; case LLAMA_FTYPE_MOSTLY_Q4_K_M: return "Q4_K - Medium"; case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "Q5_K - Small"; case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "Q5_K - Medium"; @@ -15870,6 +15871,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n 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; + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) { if (qs.model.hparams.n_vocab >= 127999 && (qs.model.type == MODEL_8B || qs.model.type == MODEL_70B)) new_type = GGML_TYPE_Q6_K; @@ -15964,6 +15966,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) { new_type = GGML_TYPE_Q5_K; } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) { + new_type = GGML_TYPE_Q5_K; + } else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || ftype == LLAMA_FTYPE_MOSTLY_Q5_0) && qs.has_imatrix && i_layer < n_layer/8) { // Guard against craziness in the first few ffn_down layers that can happen even with imatrix for Q4_0/Q5_0. @@ -15983,7 +15988,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S || ftype == LLAMA_FTYPE_MOSTLY_IQ3_M || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_K || - ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K || + ftype == LLAMA_FTYPE_MOSTLY_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) { new_type = GGML_TYPE_Q5_K; } @@ -16051,8 +16056,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_S || new_type == GGML_TYPE_IQ3_XXS || new_type == GGML_TYPE_IQ1_S || new_type == GGML_TYPE_IQ3_S || new_type == GGML_TYPE_IQ1_M || new_type == GGML_TYPE_IQ4_K || new_type == GGML_TYPE_IQ2_K || - new_type == GGML_TYPE_IQ5_K || new_type == GGML_TYPE_IQ3_K || - new_type == GGML_TYPE_IQ6_K || new_type == GGML_TYPE_IQ4_KS || + 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) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; @@ -16085,8 +16090,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_KSS: case GGML_TYPE_IQ4_KS: + 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_Q4_K_R4: case GGML_TYPE_Q4_K: new_type = GGML_TYPE_Q5_0; break; case GGML_TYPE_IQ5_K: case GGML_TYPE_Q5_K: new_type = GGML_TYPE_Q6_0; break; @@ -16179,6 +16186,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_Q3_K_L: default_type = GGML_TYPE_Q3_K; break; case LLAMA_FTYPE_MOSTLY_Q4_K_S: case LLAMA_FTYPE_MOSTLY_Q4_K_M: default_type = GGML_TYPE_Q4_K; break; + case LLAMA_FTYPE_MOSTLY_Q4_K_R4: default_type = GGML_TYPE_Q4_K_R4; break; case LLAMA_FTYPE_MOSTLY_Q5_K_S: case LLAMA_FTYPE_MOSTLY_Q5_K_M: default_type = GGML_TYPE_Q5_K; break; case LLAMA_FTYPE_MOSTLY_Q6_K: default_type = GGML_TYPE_Q6_K; break; @@ -16581,6 +16589,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q8_0; else chunk_size_multiplier = 4; } + else if (new_type == GGML_TYPE_Q4_K_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_K; + else chunk_size_multiplier = 4; + } else if (new_type == GGML_TYPE_IQ2_BN_R4) { if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_IQ2_BN; else chunk_size_multiplier = 4;