SubTrack++ : Gradient Subspace Tracking for Scalable LLM Training
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866918170493714432 |
|---|---|
| author | Rajabi, Sahar Nonta, Nayeema Rambhatla, Sirisha |
| author_facet | Rajabi, Sahar Nonta, Nayeema Rambhatla, Sirisha |
| contents | Training large language models (LLMs) is highly resource-intensive due to their massive number of parameters and the overhead of optimizer states. While recent work has aimed to reduce memory consumption, such efforts often entail trade-offs among memory efficiency, training time, and model performance. Yet, true democratization of LLMs requires simultaneous progress across all three dimensions. To this end, we propose SubTrack++ that leverages Grassmannian gradient subspace tracking combined with projection-aware optimizers, enabling Adam's internal statistics to adapt to subspace changes. Additionally, employing recovery scaling, a technique that restores information lost through low-rank projections, further enhances model performance. Our method demonstrates SOTA convergence by exploiting Grassmannian geometry, reducing pre-training wall-time by up to 65% and fine-tuning time by 36% compared to existing SOTA methods, while maintaining the same memory footprint. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_01586 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | SubTrack++ : Gradient Subspace Tracking for Scalable LLM Training Rajabi, Sahar Nonta, Nayeema Rambhatla, Sirisha Machine Learning Training large language models (LLMs) is highly resource-intensive due to their massive number of parameters and the overhead of optimizer states. While recent work has aimed to reduce memory consumption, such efforts often entail trade-offs among memory efficiency, training time, and model performance. Yet, true democratization of LLMs requires simultaneous progress across all three dimensions. To this end, we propose SubTrack++ that leverages Grassmannian gradient subspace tracking combined with projection-aware optimizers, enabling Adam's internal statistics to adapt to subspace changes. Additionally, employing recovery scaling, a technique that restores information lost through low-rank projections, further enhances model performance. Our method demonstrates SOTA convergence by exploiting Grassmannian geometry, reducing pre-training wall-time by up to 65% and fine-tuning time by 36% compared to existing SOTA methods, while maintaining the same memory footprint. |
| title | SubTrack++ : Gradient Subspace Tracking for Scalable LLM Training |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2502.01586 |