SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2402.10193 · NeurIPS · 2024

BitDelta: Your Fine-Tune May Only Be Worth One Bit

Song Han, Tianle Cai, Kai Li, Tri Dao, Guangxuan Xiao, James Liu, Jason D. Lee

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 9 functions out of this paper's own repositories and ran 6 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
FasterDecoding/BitDelta canonical 6 of 9
FunctionStatusWhere it lives
find_corr_stddev Ran FasterDecoding/BitDelta/bitdelta/misc.py
code served (permissive licence) · get_code("788def677abe4a95")
get_dataloader Ran FasterDecoding/BitDelta/bitdelta/data.py
code served (permissive licence) · get_code("c39b5bb17824dd6e")
pack Ran FasterDecoding/BitDelta/bitdelta/binary_gemm_kernel.py
code served (permissive licence) · get_code("3857e2268fd2be5c")
parse_device Ran FasterDecoding/BitDelta/bitdelta/utils.py
code served (permissive licence) · get_code("9dcbdcf4fe815b37")
parse_dict Ran FasterDecoding/BitDelta/bitdelta/utils.py
code served (permissive licence) · get_code("c85ef5aa3bf5041e")
unpack Ran FasterDecoding/BitDelta/bitdelta/binary_gemm_kernel.py
code served (permissive licence) · get_code("7b1afaa65397beea")
dequantize_model Not yet run FasterDecoding/BitDelta/bitdelta/misc.py
code served (permissive licence) · get_code("ca14e14c001835fe")
get_dataset Not yet run FasterDecoding/BitDelta/bitdelta/data.py
code served (permissive licence) · get_code("3ab4b98edc5243cf")
get_model Not yet run FasterDecoding/BitDelta/bitdelta/utils.py
code served (permissive licence) · get_code("f08bdafbdaf3bee1")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Large Language Models (LLMs) are typically trained in two phases: pre-training on large internet-scale datasets, and fine-tuning for downstream tasks. Given the higher computational demand of pre-training, it is intuitive to assume that finetuning adds less new information to the model, and is thus more compressible. We explore this assumption by decomposing the weights of fine-tuned models into their pre-trained components and an additional delta. We introduce a simple postfine-tuning method, BitDelta, which successfully quantizes this delta down to 1 bit without compromising performance. This interesting finding not only highlights the potential redundancy of information added during fine-tuning, but also has significant implications for the multi-tenant serving and multi-tenant storage of fine-tuned models. By enabling the use of a single high-precision base model accompanied by multiple 1-bit deltas, BitDelta dramatically reduces GPU memory requirements by more than 10×, thus reducing per-user generation latency by more than 10× in multi-tenant settings. We validate BitDelta through experiments across Llama-2, Mistral and MPT model families, and on models up to 70B parameters, showcasing minimal performance degradation in all tested settings.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2402.10193")
get_code_for_paper("2402.10193")
have("2402.10193")

Connect an agent — have() is free.