diff options
| author | Iwan Kawrakow <iwan.kawrakow@gmail.com> | 2024-08-01 16:08:32 +0300 |
|---|---|---|
| committer | Kawrakow <48489457+ikawrakow@users.noreply.github.com> | 2024-08-09 16:00:31 +0200 |
| commit | cfb0410067be051dd8ea76c08280a3b04d5a5188 (patch) | |
| tree | 64ffc06e18f75a8cfc4d04b98875f68cd7fa99d4 /ggml/src/ggml.c | |
| parent | a9f302ebe2373321c12b01d8760904901aa064a4 (diff) | |
iq6_k: WIP (nothing works)
Diffstat (limited to 'ggml/src/ggml.c')
| -rw-r--r-- | ggml/src/ggml.c | 21 |
1 files changed, 21 insertions, 0 deletions
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 5c817030..b5fdb96d 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1040,6 +1040,18 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .vec_dot_type = GGML_TYPE_Q8_K, .nrows = 1, }, + [GGML_TYPE_IQ6_K] = { + .type_name = "iq6_k", + .blck_size = QK_K, + .type_size = sizeof(block_iq6_k), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_iq6_k, + .from_float = quantize_row_iq6_k, + .from_float_ref = (ggml_from_float_t)quantize_row_iq6_k_ref, + .vec_dot = vec_dot_iq6_k_q8_k, + .vec_dot_type = GGML_TYPE_Q8_K, + .nrows = 1, + }, }; // For internal test use @@ -3394,6 +3406,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_IQ3_K: wtype = GGML_TYPE_IQ3_K; break; case GGML_FTYPE_MOSTLY_IQ4_K: wtype = GGML_TYPE_IQ4_K; break; case GGML_FTYPE_MOSTLY_IQ5_K: wtype = GGML_TYPE_IQ5_K; break; + case GGML_FTYPE_MOSTLY_IQ6_K: wtype = GGML_TYPE_IQ6_K; break; case GGML_FTYPE_MOSTLY_IQ3_S: wtype = GGML_TYPE_IQ3_S; break; case GGML_FTYPE_MOSTLY_IQ2_S: wtype = GGML_TYPE_IQ2_S; break; case GGML_FTYPE_MOSTLY_Q4_0_4_4: wtype = GGML_TYPE_Q4_0_4_4; break; @@ -9648,6 +9661,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: case GGML_TYPE_IQ5_K: + case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: case GGML_TYPE_Q4_0_4_4: @@ -10033,6 +10047,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: case GGML_TYPE_IQ5_K: + case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: case GGML_TYPE_Q4_0_4_4: @@ -10168,6 +10183,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: case GGML_TYPE_IQ5_K: + case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: case GGML_TYPE_Q4_0_4_4: @@ -13092,6 +13108,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: case GGML_TYPE_IQ5_K: + case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: case GGML_TYPE_Q4_0_4_4: @@ -13287,6 +13304,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: case GGML_TYPE_IQ5_K: + case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: case GGML_TYPE_Q4_0_4_4: @@ -13556,6 +13574,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: case GGML_TYPE_IQ5_K: + case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: case GGML_TYPE_Q4_0_4_4: @@ -14152,6 +14171,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_IQ3_K: case GGML_TYPE_IQ4_K: case GGML_TYPE_IQ5_K: + case GGML_TYPE_IQ6_K: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: case GGML_TYPE_Q8_K: @@ -20892,6 +20912,7 @@ size_t ggml_quantize_chunk( 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_IQ4_K: result = quantize_iq4_k (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ5_K: result = quantize_iq5_k (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ6_K: result = quantize_iq6_k (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q4_0_4_8: result = quantize_q4_0_4x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_Q4_0_8_8: result = quantize_q4_0_8x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; |