Deep Learning Recommendation Model

Meta’s February 2024 research presents DLRMs as an established approach that had underpinned leading recommendation methods for about a decade2. Against this established approach, Meta’s research compares generative recommenders, which use sequences of people’s past interactions and other information to find or order content2.

To explain how DLRMs handle category information, Meta’s May 2023 research describes embedding tables: lookups that turn identifiers, such as an ID identifying a particular YouTube video, into lists of numbers3. Meta’s research identifies these tables as a storage and processing bottleneck, reporting that tables in Meta DLRMs require very large amounts of memory3. The scale of this problem matters because Meta’s research reports that DLRMs account for more than 50% of Meta’s computing demand for training and 80% of its demand for producing predictions; those figures concern Meta, not Instagram specifically3. Meta’s research therefore investigates predicting the computing and data-transfer costs of dividing tables among devices, then searching for a balanced arrangement that reduces training bottlenecks3.

Meta’s July 2025 research addresses a different part of DLRM training: how examples are grouped, rather than where model tables are stored13. Meta’s research describes request-only optimization, which groups all items shown in response to one user request instead of treating each item separately1. Meta’s research uses the resulting reduction in repeated storage and computation to support DLRMs with a much more complex request-level component, while largely offsetting the extra computing cost that this complexity would otherwise require1.

For the stage after training, Meta’s November 2020 research examines inference for large DLRM-like models: using a model’s learned patterns to produce recommendation predictions4. Meta’s research studies dividing these models across multiple servers so they fit in memory while still returning recommendations within tight time limits, without requiring specially configured hardware4. To evaluate the extra work involved in this arrangement, Meta’s research tests three models called DRM1, DRM2 and DRM3 with different inputs and embedding-table characteristics4. Meta’s research explicitly says these are early candidates drawn from many possible, evolving models, not standard benchmark models intended to serve as common tests4.