diff options
Diffstat (limited to 'llama.cpp')
| -rw-r--r-- | llama.cpp | 10 |
1 files changed, 4 insertions, 6 deletions
@@ -2377,12 +2377,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s printf("size = %8.3f MB\n", tensor.size/1024.0/1024.0); } else { new_type = quantized_type; - // TODO: temporary disabled until Metal / OpenCL support is available - // ref: https://github.com/ggerganov/llama.cpp/issues/1711 - //if (tensor.name == "output.weight") { - // new_type = GGML_TYPE_Q6_K; - //} - if (tensor.name.find("attention.wv.weight") != std::string::npos) { + if (tensor.name == "output.weight") { + new_type = GGML_TYPE_Q6_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 = 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) && |