ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-14 15:57:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq3_kt WIP: slowly improving

Browse Source 
PPL(LLaMA-3.1-8B-Instruct, 8192) is now 6.7689 after shrinking
by 0.015 bpw by using iq4_k instead of q5_k for attn_v.
This commit is contained in:
Iwan Kawrakow
2024-11-09 11:42:14 +02:00
parent dfcc8a9cf3
commit 8f0d075f5e
2 changed files with 12 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llama.cpp
View File

@@ -15819,7 +15819,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
: !qs.has_imatrix ? GGML_TYPE_IQ3_S : GGML_TYPE_IQ3_XXS;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KT) {
new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ3_K
new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_IQ4_K : qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ3_K
: !qs.has_imatrix ? GGML_TYPE_IQ3_K : GGML_TYPE_IQ3_KT;
}
else if ((ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S) && qs.model.hparams.n_gqa() >= 2) {