We introduce an innovative and out-of-box KV cache compression method, SnapKV.

SnapKV-optimized models are all under models file, which could be directly imported and used the same like baseline models. For example:

from models.modeling_mistral import MistralForCausalLM as SnapKVMistralForCausalLM
model = SnapKVMistralForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True,
    device_map="auto",
    use_flash_attention_2=True
)
tokenizer = transformers.AutoTokenizer.from_pretrained(
    model_name,
    padding_side="right",
    use_fast=False,
)

SnapKV can be easily integrate with other models. You can follow the comment marked with [SnapKV] in existing models to constrcut your own models. The detailed algorithm of SnapKV is in snapkv_utils.py

The observations and motivations behind SnapKV could be found at observation folder