mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2025-09-16 20:09:41 +00:00
skip unquantizable clip layers
This commit is contained in:
parent
b8883e254a
commit
89a3742ded
1 changed files with 11 additions and 0 deletions
|
@ -4149,6 +4149,17 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
|
|||
{
|
||||
quantize = false;
|
||||
}
|
||||
for(int d=0;d<ggml_n_dims(cur);++d)
|
||||
{
|
||||
const int64_t blck_size = ggml_blck_size(type);
|
||||
if(d==0 && cur->ne[d] % blck_size != 0)
|
||||
{
|
||||
printf("\nSkipping %s because %d is not divisible by %d\n",name.c_str(),cur->ne[d],blck_size);
|
||||
quantize = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
// //temp fix for gemma3
|
||||
// if(name.find("ffn_up.weight") != std::string::npos)
|
||||
// {
|
||||
|
|
Loading…
Add table
Add a link
Reference in a new issue