vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

As an alternative, to avoid failing on Metal due to lack of Q8_0 support, instead quantize tok_embeddings.weight to Q4_0 and retain output.weight as F16. This results in a net gain of about 55mb for a 7B model compared to previous approach, but should minimize adverse impact to model quality.

Browse source
This commit is contained in:
Concedo 2023-07-10 23:22:45 +08:00
parent 048dca9809
commit fd9a2fdfe2
1 changed files with 9 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

11
llama.cpp
View file

@ -2435,7 +2435,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
int ny = tensor.ne.at(1);
if (nx % QK_K != 0 || ny % QK_K != 0) {
fprintf(stderr, "\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K);
fprintf(stderr, "Q8_0 will be used for this tensor instead.\n");
convert_incompatible_tensor = true;
}
}
@ -2465,7 +2464,15 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
}
if (convert_incompatible_tensor) {
new_type = GGML_TYPE_Q8_0; //fall back to Q8_0 instead of just failing.
if (tensor.name == "output.weight") {
new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing.
fprintf(stderr, "F16 will be used for this tensor instead.\n");
} else if (tensor.name == "tok_embeddings.weight") {
new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing.
fprintf(stderr, "Q4_0 will be used for this tensor instead.\n");
} else {
throw std::runtime_error("Unsupported tensor size encountered\n");
}
}
#endif