In the quantization literature they always ignore the token embedding and output tensors (they leave them as f16). But when llama.cpp loads a model, it prints a bits-per-weight (bpw) value that is basically total file size on disk / total number of parameters. As this includes the output tensor, which is almost always quantized with more bpw, this makes the i- and k-quants appear not competitive.
So, this PR adds an additional print out that tells us the model size excluding token_embd.weight and output.weight, and the corresponding bpw. Here is an example from LLaMA-3.1-8B-Instruct quantized with IQ2_XS:
...
llm_load_print_meta: model type = 8B
llm_load_print_meta: model ftype = IQ2_XS - 2.3125 bpw
llm_load_print_meta: model params = 8.030 B
llm_load_print_meta: model size = 3.880 GiB (4.150 BPW)
llm_load_print_meta: repeating layers = 1.923 GiB (2.366 BPW, 6.980 B parameters)
llm_load_print_meta: general.name = Meta Llama 3.1 8B Instruct
...
I also added one extra digit (two decimal places is a bit too little for bpw values).
In the quantization literature they always ignore the token embedding and output tensors (they leave them as
f16). But whenllama.cpploads a model, it prints a bits-per-weight (bpw) value that is basicallytotal file size on disk / total number of parameters. As this includes the output tensor, which is almost always quantized with more bpw, this makes the i- and k-quants appear not competitive.So, this PR adds an additional print out that tells us the model size excluding
token_embd.weightandoutput.weight, and the corresponding bpw. Here is an example from LLaMA-3.1-8B-Instruct quantized withIQ2_XS:I also added one extra digit (two decimal places is a bit too little for bpw values).