thunlp / InfLLM

The code of our paper "InfLLM: Unveiling the Intrinsic Capacity of LLMs for Understanding Extremely Long Sequences with Training-Free Memory"
MIT License
269 stars 21 forks source link

longbench #35

Closed Michelleable closed 4 months ago

Michelleable commented 4 months ago

Hi, there are some differences between the longbench indicator I ran and the paper. Is there any wrong setting in my experiment? Thanks.

model: type: inf-llm path: Mistral-7B-Instruct-v0.2 fattn: false block_size: 128 base: 1000000 distance_scale: 1.0 n_init: 128
n_local: 4096 topk: 16
repr_topk: 4 max_cached_block: 32 exc_block_size: 512 async_global_stream: true cache_strategy: lru faiss: false

max_len: 2147483647 truncation: suffix chunk_size: 8192

conv_type: mistral-inst