Skip to content

Quantization improvements for k_quants #2707

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 12 commits into from
Aug 22, 2023
Merged

Quantization improvements for k_quants #2707

Changes from 1 commit
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
f26f9ef
Improve LLaMA-2 2-, 3- and 4-bit quantization
Aug 13, 2023
77aea72
Minor 4-bit quantization improvement
Aug 13, 2023
ec9cb75
Some more fine tuning
Aug 13, 2023
4f8dcb1
Adding make_qkx2_quants
Aug 14, 2023
e9f1340
Another minor improvement
Aug 14, 2023
1c1f985
Q2_K improvement
Aug 14, 2023
404e43c
Iterating
Aug 16, 2023
9f78d4c
Revert Q5_K back to make_qkx1_quants
Aug 20, 2023
e2af308
Better Q6_K
Aug 21, 2023
b706339
make_qkx2_quants is better for Q5_K after all
Aug 22, 2023
35a0b97
Fix after rebasing on master
Aug 22, 2023
fdf73db
Fix for changed tensor names
Aug 22, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Next Next commit
Improve LLaMA-2 2-, 3- and 4-bit quantization 
* Q3_K_S: use Q5_K for 1st 2 layers of attention.wv and feed_forward.w2
* Q4_K_S: use Q6_K for 1st 2 layers of attention.wv and feed_forward.w2
* Q2_K and Q3_K_M: use Q5_K instead of Q4_K for 1st 2 layers of
  attention.wv and feed_forward.w2

This leads to a slight model sized increase as follows:
Q2_K  : 2.684G vs 2.670G
Q3_K_S: 2.775G vs 2.745G
Q3_K_M: 3.071G vs 3.057G
Q4_K_S: 3.592G vs 3.563G

LLaMA-2 PPL for context 512 changes as follows:
Q2_K  : 6.6691 vs 6.8201
Q3_K_S: 6.2129 vs 6.2584
Q3_K_M: 6.0387 vs 6.1371
Q4_K_S: 5.9138 vs 6.0041

There are improvements for LLaMA-1 as well, but they are
way smaller than the above.
  • Loading branch information
Iwan Kawrakow committed Aug 22, 2023
commit f26f9ef42c0e6d105b79c89447c2cd9d2c9c4763
13 changes: 12 additions & 1 deletion llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3718,18 +3718,29 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
}
} else if (name.find("attn_v.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
} else if (tensor.name.find("attention.wv.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
new_type = i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
use_more_bits(i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_attention_wv < 2) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S && i_attention_wv < 2) new_type = GGML_TYPE_Q5_K;
else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
(i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
++i_attention_wv;
} else if (name.find("ffn_down.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
} else if (tensor.name.find("feed_forward.w2.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) {
new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
//else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_feed_forward_w2 < n_feed_forward_w2/8) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_feed_forward_w2 < 2) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S && i_feed_forward_w2 < 2) new_type = GGML_TYPE_Q5_K;
++i_feed_forward_w2;
} else if (name.find("attn_output.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
Expand Down