Rockmate: an Efficient, Fast, Automatic and Generic Tool for Re-materialization in PyTorch - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Conference Papers Year : 2023

Rockmate: an Efficient, Fast, Automatic and Generic Tool for Re-materialization in PyTorch

Abstract

We propose Rockmate to control the memory requirements when training PyTorch DNN models. Rockmate is an automatic tool that starts from the model code and generates an equivalent model, using a predefined amount of memory for activations, at the cost of a few re-computations. Rockmate automatically detects the structure of computational and data dependencies and rewrites the initial model as a sequence of complex blocks. We show that such a structure is widespread and can be found in many models in the literature (Transformer based models, ResNet, RegNets,...). This structure allows us to solve the problem in a fast and efficient way, using an adaptation of Checkmate (too slow on the whole model but general) at the level of individual blocks and an adaptation of Rotor (fast but limited to sequential models) at the level of the sequence itself. We show through experiments on many models that Rockmate is as fast as Rotor and as efficient as Checkmate, and that it allows in many cases to obtain a significantly lower memory consumption for activations (by a factor of 2 to 5) for a rather negligible overhead (of the order of 10% to 20%). Rockmate is open source and available at https://github.com/topal-team/rockmate.
Fichier principal
Vignette du fichier
main.pdf (621.99 Ko) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04095305 , version 1 (16-05-2023)
hal-04095305 , version 2 (01-06-2023)
hal-04095305 , version 3 (06-06-2023)
hal-04095305 , version 4 (14-06-2023)
hal-04095305 , version 5 (26-06-2023)

Identifiers

  • HAL Id : hal-04095305 , version 5

Cite

Xunyi Zhao, Théotime Le Hellard, Lionel Eyraud-Dubois, Julia Gusak, Olivier Beaumont. Rockmate: an Efficient, Fast, Automatic and Generic Tool for Re-materialization in PyTorch. ICML 2023 - 40 th International Conference on Machine Learning, Jul 2023, Honolulu (HI), United States. ⟨hal-04095305v5⟩
184 View
128 Download

Share

Gmail Mastodon Facebook X LinkedIn More