llama : register the DT3 file type

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 23:33:14 +02:00
parent 698f37f40b
commit 693eb7d719
4 changed files with 8 additions and 2 deletions
+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;