ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 11:21:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

bitnet: add 2 bpw quantization

Browse Source 
The scalar dot product already chieves 37 t/s for TG!
This commit is contained in:
Iwan Kawrakow
2024-06-17 18:41:30 +03:00
parent 765622ff8f
commit f6863cfa1b
10 changed files with 216 additions and 184 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
ggml.c
View File

@@ -874,6 +874,18 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.vec_dot_type = GGML_TYPE_Q8_K64,
.nrows = 1,
},
[GGML_TYPE_IQ2_BN] = {
.type_name = "iq2_bn",
.blck_size = QK_IQ1BN,
.type_size = sizeof(block_iq2_bn),
.is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq2_bn,
.from_float = quantize_row_iq2_bn,
.from_float_reference = (ggml_from_float_t)quantize_row_iq2_bn_reference,
.vec_dot = ggml_vec_dot_iq2_bn_q8_K64,
.vec_dot_type = GGML_TYPE_Q8_K64,
.nrows = 1,
},
[GGML_TYPE_IQ4_NL] = {
.type_name = "iq4_nl",
.blck_size = QK4_NL,
@@ -3206,6 +3218,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
case GGML_FTYPE_MOSTLY_IQ1_S: wtype = GGML_TYPE_IQ1_S; break;
case GGML_FTYPE_MOSTLY_IQ1_M: wtype = GGML_TYPE_IQ1_M; break;
case GGML_FTYPE_MOSTLY_IQ1_BN: wtype = GGML_TYPE_IQ1_BN; break;
case GGML_FTYPE_MOSTLY_IQ2_BN: wtype = GGML_TYPE_IQ2_BN; break;
case GGML_FTYPE_MOSTLY_IQ4_NL: wtype = GGML_TYPE_IQ4_NL; break;
case GGML_FTYPE_MOSTLY_IQ4_XS: wtype = GGML_TYPE_IQ4_XS; break;
case GGML_FTYPE_MOSTLY_IQ3_S: wtype = GGML_TYPE_IQ3_S; break;
@@ -9500,6 +9513,7 @@ static void ggml_compute_forward_add(
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ1_BN:
case GGML_TYPE_IQ2_BN:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
@@ -9900,6 +9914,7 @@ static void ggml_compute_forward_add1(
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ1_BN:
case GGML_TYPE_IQ2_BN:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
@@ -10030,6 +10045,7 @@ static void ggml_compute_forward_acc(
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ1_BN:
case GGML_TYPE_IQ2_BN:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
@@ -13003,6 +13019,7 @@ static void ggml_compute_forward_out_prod(
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ1_BN:
case GGML_TYPE_IQ2_BN:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
@@ -13197,6 +13214,7 @@ static void ggml_compute_forward_set(
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ1_BN:
case GGML_TYPE_IQ2_BN:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
@@ -13473,6 +13491,7 @@ static void ggml_compute_forward_get_rows(
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ1_BN:
case GGML_TYPE_IQ2_BN:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
@@ -14087,6 +14106,7 @@ static void ggml_compute_forward_clamp(
case GGML_TYPE_IQ1_S:
case GGML_TYPE_IQ1_M:
case GGML_TYPE_IQ1_BN:
case GGML_TYPE_IQ2_BN:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_IQ4_XS:
case GGML_TYPE_IQ3_S:
@@ -21082,6 +21102,7 @@ void ggml_quantize_init(enum ggml_type type) {
case GGML_TYPE_IQ1_M: iq2xs_init_impl(type); break;
case GGML_TYPE_IQ3_XXS: iq3xs_init_impl(256); break;
case GGML_TYPE_IQ3_S: iq3xs_init_impl(512); break;
case GGML_TYPE_IQ2_BN:
case GGML_TYPE_IQ1_BN: iq1bn_init_impl(); break;
default: // nothing
break;
@@ -21152,6 +21173,7 @@ size_t ggml_quantize_chunk(
case GGML_TYPE_IQ1_S: result = quantize_iq1_s (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_IQ1_M: result = quantize_iq1_m (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_IQ1_BN: result = quantize_iq1_bn (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_IQ2_BN: result = quantize_iq2_bn (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
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_XS: result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_F16: