diff options
Diffstat (limited to 'ggml.c')
| -rw-r--r-- | ggml.c | 26 |
1 files changed, 26 insertions, 0 deletions
@@ -573,6 +573,17 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .vec_dot = ggml_vec_dot_q6_K_q8_K, .vec_dot_type = GGML_TYPE_Q8_K, }, + [GGML_TYPE_IQ2_XXS] = { + .type_name = "iq2_xxs", + .blck_size = QK_K, + .type_size = sizeof(block_iq2_xxs), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_iq2_xxs, + .from_float = quantize_row_iq2_xxs, + .from_float_reference = (ggml_from_float_t) quantize_row_iq2_xxs_reference, + .vec_dot = ggml_vec_dot_iq2_xxs_q8_K, + .vec_dot_type = GGML_TYPE_Q8_K, + }, [GGML_TYPE_Q8_K] = { .type_name = "q8_K", .blck_size = QK_K, @@ -2111,6 +2122,7 @@ 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_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_UNKNOWN: wtype = GGML_TYPE_COUNT; break; case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break; } @@ -7436,6 +7448,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: { ggml_compute_forward_add_q_f32(params, src0, src1, dst); } break; @@ -7700,6 +7713,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: { ggml_compute_forward_add1_q_f32(params, src0, src1, dst); } break; @@ -7814,6 +7828,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: default: { GGML_ASSERT(false); @@ -10455,6 +10470,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: { ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst); } break; @@ -10629,6 +10645,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: default: { GGML_ASSERT(false); @@ -10823,6 +10840,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: { ggml_compute_forward_get_rows_q(params, src0, src1, dst); } break; @@ -11459,6 +11477,7 @@ static void ggml_compute_forward_alibi( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: case GGML_TYPE_Q8_K: case GGML_TYPE_I8: case GGML_TYPE_I16: @@ -11533,6 +11552,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: case GGML_TYPE_Q8_K: case GGML_TYPE_I8: case GGML_TYPE_I16: @@ -18648,6 +18668,12 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i block_q6_K * block = (block_q6_K*)dst + start / QK_K; result = ggml_quantize_q6_K(src + start, block, n, n, hist); } break; + case GGML_TYPE_IQ2_XXS: + { + GGML_ASSERT(start % QK_K == 0); + block_iq2_xxs * block = (block_iq2_xxs*)dst + start / QK_K; + result = ggml_quantize_iq2_xxs(src + start, block, n, n, hist); + } break; case GGML_TYPE_F16: { int elemsize = sizeof(ggml_fp16_t); |