Llama-2-7b-AQLM-2Bit-2x8-hf / configuration_llama_aqlm.py
Andrei Panferov
everythin
aebd979
raw
history blame
421 Bytes
from transformers import LlamaConfig as OrigLlamaConfig
class LlamaConfig(OrigLlamaConfig):
model_type = "llama_aqlm"
def __init__(
self,
aqlm: dict[str, int] = {
"nbits_per_codebook": 16,
"num_codebooks": 1,
"out_group_size": 8,
"in_group_size": 1,
},
**kwargs,
):
super().__init__(**kwargs)
self.aqlm = aqlm