ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-07 06:50:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Quant strategies: attn_q Q4 & attn_v Q6 for Llama 3.1 Q5_K_S (#96)

Browse Source 
* attn_q Q4 & attn_v Q6 for Llama 3.1 Q5_K_S

Pattern worth to be tested on more quants and on L3 8B.
PPL 512 = -0.024 for 70b ; - 0.005 for 8b
Size = - 640MiB for 70b ; - 64MiB for 8b

70b Q5_K_S now beats Q5_K_M by -0.012 ppl

I suspect that it goes for L3 as well, which was quite insensitive to attn_q quantization.

* indent
This commit is contained in:
Nexes the Elder
2024-10-19 17:24:43 +02:00
committed by GitHub
parent 7b886ae3d8
commit a077f09bcb
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/llama.cpp
View File

@@ -15756,6 +15756,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) {
if (qs.model.hparams.n_vocab >= 127999 && (qs.model.type == MODEL_8B || qs.model.type == MODEL_70B))
new_type = GGML_TYPE_Q6_K;
}
if (qs.model.type == MODEL_70B) {
// In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
// 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
@@ -15796,6 +15800,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ2_S;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) {
if (qs.model.hparams.n_vocab >= 127999 && (qs.model.type == MODEL_8B || qs.model.type == MODEL_70B))
new_type = GGML_TYPE_Q4_K;
}
} else if (name.find("ffn_down") != std::string::npos) {
auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str());
int i_layer = info.first, n_layer = info.second;