diff options
| -rw-r--r-- | examples/quantize/quantize.cpp | 1 | ||||
| -rw-r--r-- | ggml/include/ggml.h | 2 | ||||
| -rw-r--r-- | ggml/src/ggml-quants.c | 1 | ||||
| -rw-r--r-- | ggml/src/ggml.c | 26 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_mul_mat.cpp | 194 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.cpp | 78 | ||||
| -rw-r--r-- | ggml/src/iqk/iqk_quantize.h | 6 | ||||
| -rw-r--r-- | include/llama.h | 1 | ||||
| -rw-r--r-- | src/llama.cpp | 12 |
9 files changed, 320 insertions, 1 deletions
diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 9e0dc3cf..6cac41a2 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -42,6 +42,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = { { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", }, { "IQ4_NL_X4",LLAMA_FTYPE_MOSTLY_IQ4_NL_X4," 4.50 bpw non-linear quantization", }, { "Q4_0_R4", LLAMA_FTYPE_MOSTLY_Q4_0_R4, " 4.50 bpw non-linear quantization", }, + { "Q8_0_R4", LLAMA_FTYPE_MOSTLY_Q8_0_R4, " 8.50 bpw non-linear quantization", }, { "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.25 bpw non-linear quantization", }, { "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", }, diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 1a46881a..2358fb76 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -408,6 +408,7 @@ extern "C" { GGML_TYPE_IQ4_KSS = 146, GGML_TYPE_Q4_0_R4 = 202, + GGML_TYPE_Q8_0_R4 = 208, GGML_TYPE_IQ4_NL_X4 = 220, GGML_TYPE_COUNT, }; @@ -469,6 +470,7 @@ extern "C" { GGML_FTYPE_MOSTLY_IQ4_KSS = 139, // except 1d tensors // GGML_FTYPE_MOSTLY_Q4_0_R4 = 202, // except 1d tensors + GGML_FTYPE_MOSTLY_Q8_0_R4 = 207, // except 1d tensors GGML_FTYPE_MOSTLY_IQ4_NL_X4 = 219, // except 1d tensors }; diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index dd43e1c1..3140fc19 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -15198,6 +15198,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte case GGML_TYPE_IQ4_KSS: break; case GGML_TYPE_IQ4_NL_X4: break; case GGML_TYPE_Q4_0_R4: break; + case GGML_TYPE_Q8_0_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 dfe1017b..fd65ae67 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1279,6 +1279,23 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .nrows = 1, .row_meta_size = 0, }, + [GGML_TYPE_Q8_0_R4] = { + .type_name = "q8_0_r4", + .blck_size = QK8_0, + .type_size = sizeof(block_q8_0), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_q8_0_r4, + .from_float = quantize_row_q8_0_r4, + .from_float_ref = (ggml_from_float_t)quantize_row_q8_0_r4_ref, + .vec_dot = vec_dot_q8_0_r4_q8_0, +#if GGML_USE_IQK_MULMAT && defined __AVX2__ + .vec_dot_type = GGML_TYPE_Q8_1, +#else + .vec_dot_type = GGML_TYPE_Q8_0, +#endif + .nrows = 1, + .row_meta_size = 0, + }, }; // For internal test use @@ -3939,6 +3956,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_IQ4_NL: wtype = GGML_TYPE_IQ4_NL; break; case GGML_FTYPE_MOSTLY_IQ4_NL_X4: wtype = GGML_TYPE_IQ4_NL_X4;break; case GGML_FTYPE_MOSTLY_Q4_0_R4: wtype = GGML_TYPE_Q4_0_R4; break; + case GGML_FTYPE_MOSTLY_Q8_0_R4: wtype = GGML_TYPE_Q8_0_R4; break; case GGML_FTYPE_MOSTLY_IQ4_XS: wtype = GGML_TYPE_IQ4_XS; break; case GGML_FTYPE_MOSTLY_IQ4_KS: wtype = GGML_TYPE_IQ4_KS; break; case GGML_FTYPE_MOSTLY_IQ4_KSS: wtype = GGML_TYPE_IQ4_KSS; break; @@ -10464,6 +10482,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: @@ -10908,6 +10927,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: @@ -11049,6 +11069,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: @@ -14236,6 +14257,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: @@ -14617,6 +14639,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: @@ -14892,6 +14915,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: @@ -15494,6 +15518,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_NL_X4: case GGML_TYPE_Q4_0_R4: + case GGML_TYPE_Q8_0_R4: case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_KS: case GGML_TYPE_IQ4_KSS: @@ -22322,6 +22347,7 @@ size_t ggml_quantize_chunk( case GGML_TYPE_IQ4_NL: result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_NL_X4: result = quantize_iq4_nl_x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q4_0_R4: result = quantize_q4_0_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q8_0_R4: result = quantize_q8_0_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_XS: result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; 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; diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp index 13ca6724..bbf7e379 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -2283,6 +2283,140 @@ static void mul_mat_q4_0_r4_q8_1(int n, const void * vx, size_t bx, const DataIn } #endif +#ifdef HAVE_FANCY_SIMD +template <int nrc_y> +static void mul_mat_q8_0_r4_q8_1(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { + GGML_ASSERT(nrc_x%8 == 0); + Q8<nrc_y, block_q8_1_x4> q8(info); + int nb = n / QK8_0; + GGML_ASSERT(nb%4 == 0); + if constexpr (nrc_y == 1) { + auto m127 = _mm256_set1_epi8(127); + auto m1 = _mm256_set1_epi16(1); + __m256 acc[nrc_y] = {}; + for (int ix = 0; ix < nrc_x; ix += 4) { + const block_q8_0_x4 * iq8 = (const block_q8_0_x4 *)((const char *)vx + ix*bx); + for (int ib4 = 0; ib4 < nb/4; ++ib4) { + for (int k = 0; k < 4; ++k) { + auto scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq8[4*ib4+k].d)); + auto scales = _mm256_set_m128(scales128, scales128); + auto scales_m = _mm256_mul_ps(scales, _mm256_set1_ps(-63.5f)); + auto q1 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+0), m127); + auto q2 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+1), m127); + auto q3 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+2), m127); + auto q4 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+3), m127); + for (int iy = 0; iy < nrc_y; ++iy) { + auto y = _mm256_loadu_si256((const __m256i*)q8.y[iy][ib4].qs+k); + auto sumi1 = _mm256_add_epi32(_mm256_madd_epi16(m1, _mm256_maddubs_epi16(q1, _mm256_shuffle_epi32(y, 0x00))), + _mm256_madd_epi16(m1, _mm256_maddubs_epi16(q2, _mm256_shuffle_epi32(y, 0x55)))); + auto sumi2 = _mm256_add_epi32(_mm256_madd_epi16(m1, _mm256_maddubs_epi16(q3, _mm256_shuffle_epi32(y, 0xaa))), + _mm256_madd_epi16(m1, _mm256_maddubs_epi16(q4, _mm256_shuffle_epi32(y, 0xff)))); + auto d4d8 = _mm256_mul_ps(scales, _mm256_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k]))); + acc[iy] = _mm256_fmadd_ps(d4d8, _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), acc[iy]); + acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k+4])), 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)); + info.store(ix, iy, sum); + acc[iy] = _mm256_setzero_ps(); + } + } + } else { + __m512 acc[2*nrc_y] = {}; + __m512i qx[4]; + auto m127 = _mm512_set1_epi8(127); + for (int ix = 0; ix < nrc_x; ix += 8) { + const block_q8_0_x4 * q8l = (const block_q8_0_x4 *)((const char *)vx + (ix+0)*bx); + const block_q8_0_x4 * q8h = (const block_q8_0_x4 *)((const char *)vx + (ix+4)*bx); + for (int ib4 = 0; ib4 < nb/4; ++ib4) { + for (int k = 0; k < 4; ++k) { + auto scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)q8l[4*ib4+k].d)); + auto scales1 = _mm256_set_m128(scales128, scales128); + scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)q8h[4*ib4+k].d)); + auto scales2 = _mm256_set_m128(scales128, scales128); + auto scales = _mm512_insertf32x8(_mm512_castps256_ps512(scales1), scales2, 1); + auto scales_m = _mm512_mul_ps(scales, _mm512_set1_ps(-63.5f)); + qx[0] = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)q8l[4*ib4+k].qs+0)), + _mm256_loadu_si256((const __m256i *)q8h[4*ib4+k].qs+0), 1); + qx[1] = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)q8l[4*ib4+k].qs+1)), + _mm256_loadu_si256((const __m256i *)q8h[4*ib4+k].qs+1), 1); + qx[2] = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)q8l[4*ib4+k].qs+2)), + _mm256_loadu_si256((const __m256i *)q8h[4*ib4+k].qs+2), 1); + qx[3] = _mm512_inserti32x8(_mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)q8l[4*ib4+k].qs+3)), + _mm256_loadu_si256((const __m256i *)q8h[4*ib4+k].qs+3), 1); + qx[0] = _mm512_add_epi8(qx[0], m127); + qx[1] = _mm512_add_epi8(qx[1], m127); + qx[2] = _mm512_add_epi8(qx[2], m127); + qx[3] = _mm512_add_epi8(qx[3], m127); + for (int iy = 0; iy < nrc_y; ++iy) { + auto y8 = _mm256_loadu_si256((const __m256i*)q8.y[iy][ib4].qs+k); + 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))); + auto dy = _mm512_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k])); + acc[2*iy+0] = _mm512_fmadd_ps(_mm512_mul_ps(scales, dy), _mm512_cvtepi32_ps(sumi), acc[2*iy+0]); + acc[2*iy+1] = _mm512_fmadd_ps(scales_m, _mm512_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k+4])), acc[2*iy+1]); + } + } + } + for (int iy = 0; iy < nrc_y; ++iy) { + auto sum512 = _mm512_add_ps(acc[2*iy+0], acc[2*iy+1]); + 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); + } + } + } +} +#else +template <int nrc_y> +static void mul_mat_q8_0_r4_q8_1(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_1_x4> q8(info); + auto m127 = _mm256_set1_epi8(127); + auto m1 = _mm256_set1_epi16(1); + int nb = n / QK8_0; + GGML_ASSERT(nb%4 == 0); + __m256 acc[nrc_y] = {}; + for (int ix = 0; ix < nrc_x; ix += 4) { + const block_q8_0_x4 * iq8 = (const block_q8_0_x4 *)((const char *)vx + ix*bx); + for (int ib4 = 0; ib4 < nb/4; ++ib4) { + for (int k = 0; k < 4; ++k) { + auto scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq8[4*ib4+k].d)); + auto scales = _mm256_set_m128(scales128, scales128); + auto scales_m = _mm256_mul_ps(scales, _mm256_set1_ps(-63.5f)); + auto q1 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+0), m127); + auto q2 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+1), m127); + auto q3 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+2), m127); + auto q4 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+3), m127); + for (int iy = 0; iy < nrc_y; ++iy) { + auto y = _mm256_loadu_si256((const __m256i*)q8.y[iy][ib4].qs+k); + auto sumi1 = _mm256_add_epi32(_mm256_madd_epi16(m1, _mm256_maddubs_epi16(q1, _mm256_shuffle_epi32(y, 0x00))), + _mm256_madd_epi16(m1, _mm256_maddubs_epi16(q2, _mm256_shuffle_epi32(y, 0x55)))); + auto sumi2 = _mm256_add_epi32(_mm256_madd_epi16(m1, _mm256_maddubs_epi16(q3, _mm256_shuffle_epi32(y, 0xaa))), + _mm256_madd_epi16(m1, _mm256_maddubs_epi16(q4, _mm256_shuffle_epi32(y, 0xff)))); + auto d4d8 = _mm256_mul_ps(scales, _mm256_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k]))); + acc[iy] = _mm256_fmadd_ps(d4d8, _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), acc[iy]); + acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k+4])), 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)); + info.store(ix, iy, sum); + acc[iy] = _mm256_setzero_ps(); + } + } +} +#endif + template <typename Bits> inline void multiply_add_1(int j, const Bits& bits, const __m256i * scales, const __m256i * q8, __m256i * sumi) { if (j == 0) { @@ -4264,6 +4398,18 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) { mm.funcs[7] = mul_mat_q4_0_r4_q8_1<8>; expected_typeB = GGML_TYPE_Q8_1; break; + case GGML_TYPE_Q8_0_R4: + assert (ne00 % QK4_NL == 0); + mm.funcs[0] = mul_mat_q8_0_r4_q8_1<1>; + mm.funcs[1] = mul_mat_q8_0_r4_q8_1<2>; + mm.funcs[2] = mul_mat_q8_0_r4_q8_1<3>; + mm.funcs[3] = mul_mat_q8_0_r4_q8_1<4>; + mm.funcs[4] = mul_mat_q8_0_r4_q8_1<5>; + mm.funcs[5] = mul_mat_q8_0_r4_q8_1<6>; + mm.funcs[6] = mul_mat_q8_0_r4_q8_1<7>; + mm.funcs[7] = mul_mat_q8_0_r4_q8_1<8>; + expected_typeB = GGML_TYPE_Q8_1; + break; default: return false; @@ -6805,6 +6951,43 @@ void mul_mat_q4_0_r4_q8_0(int n, const void * vx, size_t bx, const DataInfo& inf } } +template <int nrc_y> +void mul_mat_q8_0_r4_q8_0(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_0_x4> q8(info); + int nb = n / QK8_0; + GGML_ASSERT(nb%4 == 0); + float32x4_t acc[nrc_y] = {}; + for (int ix = 0; ix < nrc_x; ix += 4) { + const block_q8_0_x4 * iq8 = (const block_q8_0_x4 *)((const char *)vx + ix*bx); + for (int ib4 = 0; ib4 < nb/4; ++ib4) { + for (int k = 0; k < 4; ++k) { + auto scales = vcvt_f32_f16(vld1_f16((const float16_t *)iq8[4*ib4+k].d)); + auto qx1 = vld1q_s8_x4(iq8[4*ib4+k].qs); + auto qx2 = vld1q_s8_x4(iq8[4*ib4+k].qs+64); + for (int iy = 0; iy < nrc_y; ++iy) { + auto y = vld1q_s8_x2(q8.y[iy][ib4].qs+32*k); + auto sumi = vdupq_n_s32(0); + sumi = vdotq_laneq_s32(sumi, qx1.val[0], y.val[0], 0); + sumi = vdotq_laneq_s32(sumi, qx1.val[1], y.val[1], 0); + sumi = vdotq_laneq_s32(sumi, qx1.val[2], y.val[0], 1); + sumi = vdotq_laneq_s32(sumi, qx1.val[3], y.val[1], 1); + sumi = vdotq_laneq_s32(sumi, qx2.val[0], y.val[0], 2); + sumi = vdotq_laneq_s32(sumi, qx2.val[1], y.val[1], 2); + sumi = vdotq_laneq_s32(sumi, qx2.val[2], y.val[0], 3); + sumi = vdotq_laneq_s32(sumi, qx2.val[3], y.val[1], 3); + auto d4d8 = vmulq_f32(scales, vdupq_n_f32(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k]))); + acc[iy] = vfmaq_f32(acc[iy], d4d8, vcvtq_f32_s32(sumi)); + } + } + } + for (int iy = 0; iy < nrc_y; ++iy) { + info.store(ix, iy, acc[iy]); + acc[iy] = vdupq_n_f32(0.f); + } + } +} + template <typename Dequantizer> void MulMat::set_functions(MulMat& m) { if constexpr (std::is_same_v<Dequantizer, DequantizerQ40> || std::is_same_v<Dequantizer, DequantizerQ50> || std::is_same_v<Dequantizer, DequantizerQ80> || std::is_same_v<Dequantizer, DequantizerIQ4NL> || @@ -6996,6 +7179,17 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) { m.funcs[7] = mul_mat_q4_0_r4_q8_0<8>; expected_Btype = GGML_TYPE_Q8_0; break; + case GGML_TYPE_Q8_0_R4: + m.funcs[0] = mul_mat_q8_0_r4_q8_0<1>; + m.funcs[1] = mul_mat_q8_0_r4_q8_0<2>; + m.funcs[2] = mul_mat_q8_0_r4_q8_0<3>; + m.funcs[3] = mul_mat_q8_0_r4_q8_0<4>; + m.funcs[4] = mul_mat_q8_0_r4_q8_0<5>; + m.funcs[5] = mul_mat_q8_0_r4_q8_0<6>; + m.funcs[6] = mul_mat_q8_0_r4_q8_0<7>; + m.funcs[7] = mul_mat_q8_0_r4_q8_0<8>; + expected_Btype = GGML_TYPE_Q8_0; + break; default: return false; } diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp index b9e6ff68..811a9fe9 100644 --- a/ggml/src/iqk/iqk_quantize.cpp +++ b/ggml/src/iqk/iqk_quantize.cpp @@ -3287,3 +3287,81 @@ void vec_dot_q4_0_r4_q8_0(int n, float * s, size_t bs, const void * vx, size_t b GGML_UNUSED(by); } + +// +// ========================================= q8_0_r4 +// +void quantize_row_q8_0_r4_ref(const float * x, block_q8_0_x4 * y, int64_t k) { + // we assume we are called with 4 rows + quantize_q8_0_r4(x, (void *)y, 4, k/4, nullptr); +} + +void quantize_row_q8_0_r4(const float * x, void * y, int64_t k) { + // we assume we are called with 4 rows + quantize_q8_0_r4(x, y, 4, k/4, nullptr); +} + +static void repack_q8_0(int nrows, int n_per_row, const block_q8_0 * x, block_q8_0_x4 * y) { + GGML_ASSERT(nrows%4 == 0); + GGML_ASSERT(n_per_row%QK8_0 == 0); + int nblock = n_per_row/QK8_0; + const block_q8_0 * x4[4]; + for (int row = 0; row < nrows; row += 4) { + for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k; + for (int ib = 0; ib < nblock; ++ib) { + for (int k = 0; k < 4; ++k) y[ib].d[k] = x4[k][ib].d; + for (int l = 0; l < 4; ++l) { + for (int k = 0; k < 4; ++k) for (int i = 0; i < 4; ++i) { + y[ib].qs[32*l+4*k+i+ 0] = x4[k][ib].qs[i+4*l+ 0]; + y[ib].qs[32*l+4*k+i+16] = x4[k][ib].qs[i+4*l+16]; + } + } + } + x += 4*nblock; + y += nblock; + } +} + +size_t quantize_q8_0_r4(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) { + GGML_ASSERT(nrows%4 == 0); + auto row_size_0 = ggml_row_size(GGML_TYPE_Q8_0, n_per_row); + std::vector<char> qtmp(4*row_size_0); + char * qrow = (char *)dst; + for (int row = 0; row < nrows; row += 4) { + quantize_q8_0(src, qtmp.data(), 4, n_per_row, imatrix); + repack_q8_0(4, n_per_row, (const block_q8_0 *)qtmp.data(), (block_q8_0_x4 *)qrow); + src += 4*n_per_row; + qrow += 4*row_size_0; + } + return nrows*row_size_0; +} + +void dequantize_row_q8_0_r4(const block_q8_0_x4 * x, float * y, int64_t k) { + // we assume we are called with 4 rows + int n_per_row = k/4; + int nb = n_per_row/QK8_0; + float * yk[4]; + for (int k = 0; k < 4; ++k) yk[k] = y + k*n_per_row; + for (int ib = 0; ib < nb; ++ib) { + for (int k = 0; k < 4; ++k) { + float scale = GGML_FP16_TO_FP32(x[ib].d[k]); + for (int l = 0; l < 4; ++l) for (int i = 0; i < 4; ++i) { + yk[k][QK8_0*ib+4*l+i+ 0] = scale * x[ib].qs[QK8_0*l+4*k+i+ 0]; + yk[k][QK8_0*ib+4*l+i+16] = scale * x[ib].qs[QK8_0*l+4*k+i+16]; + } + } + } +} + +void vec_dot_q8_0_r4_q8_0(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_Q8_0_R4, vx, 0, GGML_TYPE_Q8_0, 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 98c9c010..53caed4c 100644 --- a/ggml/src/iqk/iqk_quantize.h +++ b/ggml/src/iqk/iqk_quantize.h @@ -75,6 +75,12 @@ size_t quantize_q4_0_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT ds void dequantize_row_q4_0_r4(const block_iq4_nl_x4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); void vec_dot_q4_0_r4_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 quantize_row_q8_0_r4_ref(const float * GGML_RESTRICT x, block_q8_0_x4 * GGML_RESTRICT y, int64_t k); +void quantize_row_q8_0_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); +size_t quantize_q8_0_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); +void dequantize_row_q8_0_r4(const block_q8_0_x4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); +void vec_dot_q8_0_r4_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); + #ifdef __cplusplus } #endif diff --git a/include/llama.h b/include/llama.h index fc034b3a..5e935533 100644 --- a/include/llama.h +++ b/include/llama.h @@ -181,6 +181,7 @@ extern "C" { LLAMA_FTYPE_MOSTLY_IQ4_KSS = 148, // except 1d tensors // LLAMA_FTYPE_MOSTLY_Q4_0_R4 = 202, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q8_0_R4 = 207, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ4_NL_X4 = 225, // except 1d tensors LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file diff --git a/src/llama.cpp b/src/llama.cpp index 9505f56f..89641f71 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -3851,6 +3851,7 @@ struct llama_model_loader { case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break; case GGML_TYPE_IQ4_NL_X4:ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL_X4;break; case GGML_TYPE_Q4_0_R4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_R4; break; + case GGML_TYPE_Q8_0_R4: ftype = LLAMA_FTYPE_MOSTLY_Q8_0_R4; break; case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break; case GGML_TYPE_IQ4_KS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_KS; break; case GGML_TYPE_IQ4_KSS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_KSS; break; @@ -4557,6 +4558,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ4_NL: return "IQ4_NL - 4.5 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_NL_X4:return "IQ4_NL_X4 - 4.5 bpw"; case LLAMA_FTYPE_MOSTLY_Q4_0_R4: return "Q4_0_R4 - 4.5 bpw"; + case LLAMA_FTYPE_MOSTLY_Q8_0_R4: return "Q8_0_R4 - 8.5 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_KSS: return "IQ4_KSS - 4.0 bpw"; @@ -15745,7 +15747,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KSS) && !qs.has_output) { new_type = GGML_TYPE_IQ5_K; } - else if (new_type != GGML_TYPE_Q8_0 && new_type != GGML_TYPE_IQ6_K) { + else if (new_type != GGML_TYPE_Q8_0 && new_type != GGML_TYPE_Q8_0_R4 && new_type != GGML_TYPE_IQ6_K) { new_type = GGML_TYPE_Q6_K; } } @@ -15776,6 +15778,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (new_type == GGML_TYPE_Q4_0_R4) { new_type = GGML_TYPE_Q4_0; } + else if (new_type == GGML_TYPE_Q8_0_R4) { + new_type = GGML_TYPE_Q8_0; + } } } 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 || @@ -16169,6 +16174,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ4_NL: default_type = GGML_TYPE_IQ4_NL; break; case LLAMA_FTYPE_MOSTLY_IQ4_NL_X4:default_type = GGML_TYPE_IQ4_NL_X4;break; case LLAMA_FTYPE_MOSTLY_Q4_0_R4: default_type = GGML_TYPE_Q4_0_R4; break; + case LLAMA_FTYPE_MOSTLY_Q8_0_R4: default_type = GGML_TYPE_Q8_0_R4; 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_KSS: default_type = GGML_TYPE_IQ4_KSS; break; @@ -16534,6 +16540,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0; else chunk_size_multiplier = 4; } + if (new_type == GGML_TYPE_Q8_0_R4) { + if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q8_0; + else chunk_size_multiplier = 4; + } LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type)); fflush(stdout); |