Skip to content
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

Jump to bottom

Importance matrix support for legacy quants #4969

Merged
merged 2 commits into from
Jan 16, 2024
Merged

Importance matrix support for legacy quants #4969

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6f9ec42
imatrix: adding support for legacy quants
Kawrakow Jan 16, 2024
bb9abb5
imatrix: guard Q4_0/Q5_0 against ffn_down craziness
Kawrakow Jan 16, 2024
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
Prev Previous commit
imatrix: guard Q4_0/Q5_0 against ffn_down craziness
  • Loading branch information
@Kawrakow
Kawrakow committed Jan 16, 2024
commit bb9abb5cd8d8b2517a573efc291528f47c6f6ee9
8 changes: 6 additions & 2 deletions llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8548,8 +8548,12 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && i_layer < n_layer/8) {
new_type = GGML_TYPE_Q5_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_0 && qs.has_imatrix) {
new_type = GGML_TYPE_Q4_K;
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_0 || ftype == LLAMA_FTYPE_MOSTLY_Q5_0)
&& qs.has_imatrix && i_layer < n_layer/8) {
// Guard against craziness in the first few ffn_down layers that can happen even with imatrix for Q4_0/Q5_0.
// We only do it when an imatrix is provided because a) we want to make sure that one can always get the
// same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix.
new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1;
}
++qs.i_feed_forward_w2;
} else if (name.find("attn_output.weight") != std::string::npos) {
Expand Down
Loading