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
src/llama-model-loader.cpp
+2
View File
@@ -37,6 +37,7 @@ const char * llama_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_F16: name = LLAMA_FTYPE_PREFIX "F16"; break;
case LLAMA_FTYPE_MOSTLY_BF16: name = LLAMA_FTYPE_PREFIX "BF16"; break;
case LLAMA_FTYPE_MOSTLY_Q1_0: name = LLAMA_FTYPE_PREFIX "Q1_0"; break;
case LLAMA_FTYPE_MOSTLY_Q2_0: name = LLAMA_FTYPE_PREFIX "Q2_0"; break;
case LLAMA_FTYPE_MOSTLY_Q4_0: name = LLAMA_FTYPE_PREFIX "Q4_0"; break;
case LLAMA_FTYPE_MOSTLY_Q4_1: name = LLAMA_FTYPE_PREFIX "Q4_1"; break;
case LLAMA_FTYPE_MOSTLY_Q5_0: name = LLAMA_FTYPE_PREFIX "Q5_0"; break;
@@ -767,6 +768,7 @@ llama_model_loader::llama_model_loader(
case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break;
case GGML_TYPE_NVFP4: ftype = LLAMA_FTYPE_MOSTLY_NVFP4; break;
case GGML_TYPE_Q1_0: ftype = LLAMA_FTYPE_MOSTLY_Q1_0; break;
case GGML_TYPE_Q2_0: ftype = LLAMA_FTYPE_MOSTLY_Q2_0; break;
default:
{
LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
src/llama-quant.cpp
+3 -1
View File
@@ -380,6 +380,7 @@ static ggml_type tensor_type_fallback(quantize_state_impl & qs, const ggml_tenso
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ3_S: // types on the right: block size 32
case GGML_TYPE_IQ4_XS: return_type = GGML_TYPE_IQ4_NL; break;
case GGML_TYPE_Q2_0:
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_TQ1_0:
@@ -480,7 +481,7 @@ static ggml_type llama_tensor_get_type_impl(quantize_state_impl & qs, ggml_type
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ3_S;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0) {
else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0 || ftype == LLAMA_FTYPE_MOSTLY_Q2_0) {
new_type = GGML_TYPE_Q4_K;
}
}
@@ -800,6 +801,7 @@ ggml_type llama_ftype_get_default_type(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_BF16: return GGML_TYPE_BF16;
case LLAMA_FTYPE_ALL_F32: return GGML_TYPE_F32;
case LLAMA_FTYPE_MOSTLY_Q1_0: return GGML_TYPE_Q1_0;
case LLAMA_FTYPE_MOSTLY_Q2_0: return GGML_TYPE_Q2_0;
case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: return GGML_TYPE_MXFP4;