From a63a96b5aea031f9dabf1e7c8d5ee28af170e9e7 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Tue, 10 Dec 2024 18:13:47 +0100 Subject: Q5_K_R4 (#132) * q5_k_r4: WIP * q5_k_r4: Zen4 and AVX2 We get PP-512(LLaMA-3.1-8B) = 248.3 t/s on Zen4. Q5_K_S has PP-512 = 190 t/s. * q5_k_r4: NEON We get PP-512(LLaMA-3.1-8B) = 96.1 t/s. --------- Co-authored-by: Iwan Kawrakow --- ggml/src/ggml.c | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+) (limited to 'ggml/src/ggml.c') diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index b92c2352..53c51ba6 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -914,6 +914,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .nrows = 1, .row_meta_size = 0, }, + [GGML_TYPE_Q5_K_R4] = { + .type_name = "q5_k_r4", + .blck_size = QK_K, + .type_size = sizeof(block_q5_K), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_q5_k_r4, + .from_float = quantize_row_q5_k_r4, + .from_float_ref = (ggml_from_float_t) quantize_row_q5_k_r4_ref, + .vec_dot = vec_dot_q5_k_r4_q8_k, + .vec_dot_type = GGML_TYPE_Q8_K32, + .nrows = 1, + .row_meta_size = 0, + }, [GGML_TYPE_Q6_K] = { .type_name = "q6_K", .blck_size = QK_K, @@ -4048,6 +4061,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_Q4_K: wtype = GGML_TYPE_Q4_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_Q5_K_R4: wtype = GGML_TYPE_Q5_K_R4; break; case GGML_FTYPE_MOSTLY_Q6_K: wtype = GGML_TYPE_Q6_K; break; case GGML_FTYPE_MOSTLY_Q6_K_R4: wtype = GGML_TYPE_Q6_K_R4; break; case GGML_FTYPE_MOSTLY_IQ2_XXS: wtype = GGML_TYPE_IQ2_XXS; break; @@ -10580,6 +10594,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_Q4_K: case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: + case GGML_TYPE_Q5_K_R4: case GGML_TYPE_Q6_K: case GGML_TYPE_Q6_K_R4: case GGML_TYPE_IQ2_XXS: @@ -11031,6 +11046,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_Q4_K: case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: + case GGML_TYPE_Q5_K_R4: case GGML_TYPE_Q6_K: case GGML_TYPE_Q6_K_R4: case GGML_TYPE_IQ2_XXS: @@ -11179,6 +11195,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_Q4_K: case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: + case GGML_TYPE_Q5_K_R4: case GGML_TYPE_Q6_K: case GGML_TYPE_Q6_K_R4: case GGML_TYPE_IQ2_XXS: @@ -14373,6 +14390,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_Q4_K: case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: + case GGML_TYPE_Q5_K_R4: case GGML_TYPE_Q6_K: case GGML_TYPE_Q6_K_R4: case GGML_TYPE_IQ2_XXS: @@ -14761,6 +14779,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_Q4_K: case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: + case GGML_TYPE_Q5_K_R4: case GGML_TYPE_Q6_K: case GGML_TYPE_Q6_K_R4: case GGML_TYPE_IQ2_XXS: @@ -15043,6 +15062,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_Q4_K: case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: + case GGML_TYPE_Q5_K_R4: case GGML_TYPE_Q6_K: case GGML_TYPE_Q6_K_R4: case GGML_TYPE_IQ2_XXS: @@ -15652,6 +15672,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_Q4_K: case GGML_TYPE_Q4_K_R4: case GGML_TYPE_Q5_K: + case GGML_TYPE_Q5_K_R4: case GGML_TYPE_Q6_K: case GGML_TYPE_Q6_K_R4: case GGML_TYPE_IQ2_XXS: @@ -22489,6 +22510,7 @@ size_t ggml_quantize_chunk( 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_Q5_K_R4: result = quantize_q5_k_r4(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_Q6_K_R4: result = quantize_q6_k_r4(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; -- cgit v1.2.3