Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add Q2_0 quantization: type definition and CPU backend (#24448)

Browse Source
This commit is contained in:
Pasha Khosravi
2026-07-07 12:05:47 -07:00
committed by GitHub
parent c198af4dc2
commit bec4772f6a
17 changed files with 262 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml.c
+10
View File
@@ -681,6 +681,14 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
.to_float = (ggml_to_float_t) dequantize_row_q1_0,
.from_float_ref = (ggml_from_float_t) quantize_row_q1_0_ref,
},
[GGML_TYPE_Q2_0] = {
.type_name = "q2_0",
.blck_size = QK2_0,
.type_size = sizeof(block_q2_0),
.is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q2_0,
.from_float_ref = (ggml_from_float_t) quantize_row_q2_0_ref,
},
[GGML_TYPE_Q4_0] = {
.type_name = "q4_0",
.blck_size = QK4_0,
@@ -1417,6 +1425,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
case GGML_FTYPE_MOSTLY_Q4_0: wtype = GGML_TYPE_Q4_0; break;
case GGML_FTYPE_MOSTLY_Q4_1: wtype = GGML_TYPE_Q4_1; break;
case GGML_FTYPE_MOSTLY_Q1_0: wtype = GGML_TYPE_Q1_0; break;
case GGML_FTYPE_MOSTLY_Q2_0: wtype = GGML_TYPE_Q2_0; break;
case GGML_FTYPE_MOSTLY_Q5_0: wtype = GGML_TYPE_Q5_0; break;
case GGML_FTYPE_MOSTLY_Q5_1: wtype = GGML_TYPE_Q5_1; break;
case GGML_FTYPE_MOSTLY_Q8_0: wtype = GGML_TYPE_Q8_0; break;
@@ -7744,6 +7753,7 @@ size_t ggml_quantize_chunk(
switch (type) {
case GGML_TYPE_Q1_0: result = quantize_q1_0 (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q2_0: result = quantize_q2_0 (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q4_0: result = quantize_q4_0 (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q4_1: result = quantize_q4_1 (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
case GGML_TYPE_Q5_0: result = quantize_q5_0 (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;