SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2405.18392 · NeurIPS · 2024

Scaling Laws and Compute-Optimal Training Beyond Fixed Training Durations

Martin Jaggi, Atli Kosson, Alexander Hägele, Elie Bakouch, Loubna Allal, Leandro Werra

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 13 functions out of this paper's own repositories and ran 7 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
epfml/schedules-and-scaling canonical 3 of 8
frotaur/icmlbackperp canonical 3 of 4
fabian-sp/lr-scheduling extension 1 of 1
FunctionStatusWhere it lives
calculate_slope Ran fabian-sp/lr-scheduling/reanalysis/analysis_horizon_transfer.py
code served (permissive licence) · get_code("c2122ff2161f0942")
move_to_cpu Ran epfml/schedules-and-scaling/src/logger/logger.py
code served (permissive licence) · get_code("1bd86a911c3f78a3")
multiget Ran frotaur/icmlbackperp/modules/tokenizer.py
pointer only (licence: NONE) · get_code("db9d402a23bb81a3")
precompute_freqs_cis Ran epfml/schedules-and-scaling/src/models/llama.py
code served (permissive licence) · get_code("a1d6f89d43fc42e1")
put_and_return Ran frotaur/icmlbackperp/modules/tokenizer.py
pointer only (licence: NONE) · get_code("e21c63d05fe20e51")
remove_none_vals Ran frotaur/icmlbackperp/modules/tokenizer.py
pointer only (licence: NONE) · get_code("b5644fc0a380fa78")
self_preserving_overwrite Ran epfml/schedules-and-scaling/src/logger/logger.py
code served (permissive licence) · get_code("198aea734522a443")
apply_rotary_emb Not yet run epfml/schedules-and-scaling/src/models/llama.py
code served (permissive licence) · get_code("9891174b43bd9c9a")
cos_inf_schedule Not yet run epfml/schedules-and-scaling/src/optim/utils.py
code served (permissive licence) · get_code("91d9a7a7d597054b")
get_batch Not yet run epfml/schedules-and-scaling/src/optim/utils.py
code served (permissive licence) · get_code("f9dc20d4ad58a5f9")
get_model Not yet run epfml/schedules-and-scaling/src/models/utils.py
code served (permissive licence) · get_code("843859340885ab12")
load_model Not yet run frotaur/icmlbackperp/modules/models/load_model.py
pointer only (licence: NONE) · get_code("98497683abdc8f17")
wsd_schedule Not yet run epfml/schedules-and-scaling/src/optim/utils.py
code served (permissive licence) · get_code("e8ac769090ecbe45")

Repositories linked to this paper

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

Abstract

Scale has become a main ingredient in obtaining strong machine learning models. As a result, understanding a model's scaling properties is key to effectively designing both the right training setup as well as future generations of architectures. In this work, we argue that scale and training research has been needlessly complex due to reliance on the cosine schedule, which prevents training across different lengths for the same model size. We investigate the training behavior of a direct alternative -constant learning rate and cooldowns -and find that it scales predictably and reliably similar to cosine. Additionally, we show that stochastic weight averaging yields improved performance along the training trajectory, without additional training costs, across different scales. Importantly, with these findings we demonstrate that scaling experiments can be performed with significantly reduced compute and GPU hours by utilizing fewer but reusable training runs. Our code is available at https://github.com/epfml/schedules-and-scaling/.

For agents

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

get_harvested_code_for_paper("2405.18392")
get_code_for_paper("2405.18392")
have("2405.18392")

Connect an agent — have() is free.