millaguie/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Actions 32 Packages Projects Releases Wiki Activity

llama : register the DT3 file type

Browse Source 
Adds LLAMA_FTYPE_MOSTLY_DT3 at the end of the ftype enum, the loader
name/guess mappings, the quantization fallbacks (same as the other
ternary types), and the llama-quantize table entry. The table entry
warns that the in-tree quantizer is only the reference one: DT3 models
with the measured quality are produced by the external PTQTP pipeline.
This commit is contained in:
Millaguie
2026-08-10 13:16:36 +02:00
parent d4343d0c5c
commit e1fd6e0a13
4 changed files with 8 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/llama.h
+1
View File
@@ -156,6 +156,7 @@ extern "C" {
LLAMA_FTYPE_MOSTLY_NVFP4 = 39, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q1_0 = 40, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q2_0 = 41, // except 1d tensors
LLAMA_FTYPE_MOSTLY_DT3 = 42, // except 1d tensors
LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
};
src/llama-model-loader.cpp
+2
View File
@@ -58,6 +58,7 @@ const char * llama_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_Q6_K: name = LLAMA_FTYPE_PREFIX "Q6_K"; break;
case LLAMA_FTYPE_MOSTLY_TQ1_0: name = LLAMA_FTYPE_PREFIX "TQ1_0 - 1.69 bpw ternary"; break;
case LLAMA_FTYPE_MOSTLY_TQ2_0: name = LLAMA_FTYPE_PREFIX "TQ2_0 - 2.06 bpw ternary"; break;
case LLAMA_FTYPE_MOSTLY_DT3: name = LLAMA_FTYPE_PREFIX "DT3 - 3.5 bpw dual-plane ternary"; break;
case LLAMA_FTYPE_MOSTLY_IQ2_XXS: name = LLAMA_FTYPE_PREFIX "IQ2_XXS - 2.0625 bpw"; break;
case LLAMA_FTYPE_MOSTLY_IQ2_XS: name = LLAMA_FTYPE_PREFIX "IQ2_XS - 2.3125 bpw"; break;
case LLAMA_FTYPE_MOSTLY_IQ2_S: name = LLAMA_FTYPE_PREFIX "IQ2_S - 2.5 bpw"; break;
@@ -747,6 +748,7 @@ llama_model_loader::llama_model_loader(
case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break;
case GGML_TYPE_TQ1_0: ftype = LLAMA_FTYPE_MOSTLY_TQ1_0; break;
case GGML_TYPE_TQ2_0: ftype = LLAMA_FTYPE_MOSTLY_TQ2_0; break;
case GGML_TYPE_DT3: ftype = LLAMA_FTYPE_MOSTLY_DT3; break;
case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break;
case GGML_TYPE_IQ2_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS; break;
case GGML_TYPE_IQ2_S: ftype = LLAMA_FTYPE_MOSTLY_IQ2_S; break;
src/llama-quant.cpp
+4 -2
View File
@@ -396,7 +396,8 @@ static ggml_type tensor_type_fallback(quantize_state_impl & qs, const ggml_tenso
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
case GGML_TYPE_TQ1_0:
case GGML_TYPE_TQ2_0: return_type = GGML_TYPE_Q4_0; break;
case GGML_TYPE_TQ2_0:
case GGML_TYPE_DT3: return_type = GGML_TYPE_Q4_0; break;
case GGML_TYPE_Q4_K: return_type = GGML_TYPE_Q5_0; break;
case GGML_TYPE_Q5_K: return_type = GGML_TYPE_Q5_1; break;
case GGML_TYPE_Q6_K: return_type = GGML_TYPE_Q8_0; break;
@@ -493,7 +494,7 @@ 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 || ftype == LLAMA_FTYPE_MOSTLY_Q2_0) {
else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0 || ftype == LLAMA_FTYPE_MOSTLY_Q2_0 || ftype == LLAMA_FTYPE_MOSTLY_DT3) {
new_type = GGML_TYPE_Q4_K;
}
}
@@ -831,6 +832,7 @@ ggml_type llama_ftype_get_default_type(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_Q6_K: return GGML_TYPE_Q6_K;
case LLAMA_FTYPE_MOSTLY_TQ1_0: return GGML_TYPE_TQ1_0;
case LLAMA_FTYPE_MOSTLY_TQ2_0: return GGML_TYPE_TQ2_0;
case LLAMA_FTYPE_MOSTLY_DT3: return GGML_TYPE_DT3;
case LLAMA_FTYPE_MOSTLY_IQ2_XXS: return GGML_TYPE_IQ2_XXS;
case LLAMA_FTYPE_MOSTLY_IQ2_XS: return GGML_TYPE_IQ2_XS;
case LLAMA_FTYPE_MOSTLY_IQ2_S: return GGML_TYPE_IQ2_XS;
tools/quantize/quantize.cpp
+1
View File
@@ -47,6 +47,7 @@ static const std::vector<quant_option> QUANT_OPTIONS = {
{ "IQ1_M", LLAMA_FTYPE_MOSTLY_IQ1_M, " 1.75 bpw quantization", },
{ "TQ1_0", LLAMA_FTYPE_MOSTLY_TQ1_0, " 1.69 bpw ternarization", },
{ "TQ2_0", LLAMA_FTYPE_MOSTLY_TQ2_0, " 2.06 bpw ternarization", },
{ "DT3", LLAMA_FTYPE_MOSTLY_DT3, " 3.5 bpw dual-ternary (REF quantizer only - use ternaria's PTQTP for quality)", },
{ "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.96G, +3.5199 ppl @ Llama-3-8B", },
{ "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.96G, +3.1836 ppl @ Llama-3-8B", },
{ "IQ3_XXS", LLAMA_FTYPE_MOSTLY_IQ3_XXS, " 3.06 bpw quantization", },