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

imatrix: guard even more against low-bit quantization misuse

Browse source
This commit is contained in:
Iwan Kawrakow 2024-01-12 17:41:07 +02:00
parent d5598f7ea2
commit f342143e92
1 changed files with 9 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

9
llama.cpp
View file

@ -9240,6 +9240,15 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
}
}
}
if ((new_type == GGML_TYPE_IQ2_XXS ||
new_type == GGML_TYPE_IQ2_XS ||
(new_type == GGML_TYPE_Q2_K && params->ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(tensor->name, "token_embd.weight") != 0)) && !imatrix) {
fprintf(stderr, "\n\n============================================================\n");
fprintf(stderr, "Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name);
fprintf(stderr, "The result will be garbage, so bailing out\n");
fprintf(stderr, "============================================================\n\n");
throw std::runtime_error(format("Missing importance matrix for tensor %s in a very low-bit quantization", tensor->name));
}
float * f32_data;